Wall of love for The Gen Academy

As an HR Advisor supporting organizations through their AI transformation journeys, staying close to the technical side — not just the strategic side — is something I take seriously, because it makes the advice I give sharper and more credible. That's why I recently completed Mastering Agentic AI with The Gen Academy. A few things worth noting: It is taught by people actively working/Worked with Google, NVIDIA, OpenAI, Microsoft, Fireworks.ai, Nebius, etc.— not academics. Co-founders Aishwarya Srinivasan and Arvind Narayanamurthy personally run the live sessions, and it's fully hands-on — real building happens.. Code heavy, low-code. No code and all of it...., not some theory slides. You get to learn directly from real masters and hands-on experts, and see their take on what's actually happening under the hood — the kind of insight you don't get from a textbook or a generic course from a classroom lecturer. Thank you to the entire The Gen Academy team for putting together such a credible, high-quality program. This indeed marks the new and required form of learning, that the world requires now and in the future !!

Join LinkedIn now — it’s free! Email Password (Show) Remember me By clicking Agree & Join or Continue, you agree to the LinkedIn User Agreement, Privacy Policy, and Cookie Policy. Agree & Join or Already on LinkedIn? Sign in Looking to create a page for a business? Get help

I built 𝗦𝘁𝗮𝗿𝗚𝗮𝘇𝗲𝗿, a real-time astronomy and space exploration app, as my Week 1 project in 𝗧𝗵𝗲 𝗚𝗲𝗻 𝗔𝗰𝗮𝗱𝗲𝗺𝘆’𝘀 𝗠𝗮𝘀𝘁𝗲𝗿𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 𝗔𝘂𝗴𝘂𝘀𝘁 𝗰𝗼𝗵𝗼𝗿𝘁. The idea came from planning stargazing trips to dark-sky locations like the Grand Canyon and realizing I had to use several apps to check what was visible that night, the Moon phase, weather conditions, deep-sky objects, ISS passes, and whether the conditions were actually worth the drive. I wanted to bring that into one place. StarGazer includes: • Interactive sky map and deep-sky exploration • Moon, planets, stars, and observing conditions • ISS pass information • An Analemma showing the Sun’s figure-8 path across the sky through the year • A Global Space Missions section built using CSV dataset, covering 70 years of launch history, mission success, human spaceflight, booster reuse, spending, and more It started as a stargazing app, but I also wanted it to be useful for anyone curious about astronomy and space exploration. Adding the Analemma, ISS data, and the Global Space Missions dataset helped take it in that direction. I also added Plan My Stargazing Night. It uses the app’s existing sky and weather data to give a short observing recommendation for the selected location. 🔗 Try it here: lnkd.in/g4_amKke One thing I learned while building this was how much the result depends on how clearly I explain what I want. When I was vague, I spent more time correcting things. When I gave Replit a specific requirement, screenshot, or exact issue, the iterations were much better. Brandon’s guest lecture on Replit was useful here, especially the discussion around specification, review, and orchestration. I understood those ideas much better once I started using Replit to build and refine the app. Thank you, Brandon, Aishwarya, Arvind, and The Gen Academy. If you try it, I’d be curious which section you end up exploring most. #BuildInPublic #Replit #MasteringAgenticAI #TheGenAcademy #VibeCoding #Astronomy #AI

Started Mastering Agentic AI this weekend. Six-week cohort on Maven with Aishwarya Srinivasan and Arvind Narayanamurthy, through The Gen Academy The first session was yesterday and ran past three hours. 350+ people on the call from all over. Going by the timezones in the chat, some of them were up in the middle of the night for it. Mastering agentic AI is something I've wanted to do for a long time. What actually made me sign up was Aishwarya Srinivasan. I've been following her work for a while now and the way she teaches, the way she builds, and how much she knows is what pushed me to join this cohort. I think it'll move my skills forward and help me take the next step in my career, and I'm committed to putting in the work. Six weeks, a project every week, then a capstone. A marathon, not a sprint, and I intend to finish it. Looking forward to building with AI, and I'll post what I make along the way. #AgenticAI #AI #BuildingInPublic #TheGenAcademy #Maven

A huge thank you to Brandon Middleton, Head of Education at Replit, for the guest lecture at today's The Gen Academy Bootcamp session. I know how busy someone in his position must be, so the fact that he made time for us means a lot. He walked us through building apps live on Replit — what you can actually create, and how quickly you can go from an idea to something running. Watching it happen in real time makes a point no documentation can: the gap between "I have an idea" and "it's working" is much smaller than most of us assume. Special shoutout to Aishwarya Srinivasan, Arvind Narayanamurthy for bringing Brandon in. Getting direct exposure to the people building the tools we use is a big part of why this cohort is worth it. #GenAcademy #Replit #AI #AIagents #AgenticAI #BuildInPublic #SoftwareDevelopment #DeveloperCommunity #AIEngineering #Maven

🚀 Week 2, Session 1 — RAG Deep Dive Great session today at The Gen Academy, diving into RAG, embeddings, chunking, indexing, retrieval, reranking, RAG pipelines, BM25, and Context Engineering. One key takeaway: building a strong RAG system is not just about connecting an LLM to data — it’s about retrieving the right context and giving the model exactly what it needs. Big thanks to Aishwarya Srinivasan and Arvind Narayanamurthy for making the concepts practical and easy to understand. 🙌 Excited to keep going deeper into Agentic AI and advanced RAG systems! #RAG #AgenticAI #GenerativeAI #LLM #AIEngineering #TheGenAcademy #LearningInPublic #BuildInPublic #Maven

#Week2 of the Agentic AI Cohort by The Gen Academy was packed with practical insights that are already influencing how I think about building enterprise AI systems. While I was familiar with the fundamentals of RAG, this week introduced me to several concepts that were completely new to me: 🔹 Context Rot and its impact on response quality 🔹 Lost in the Middle and how context placement affects retrieval effectiveness 🔹 Tool Loadout strategies for Agentic AI systems 🔹 GraphRAG and knowledge graph-based retrieval architectures 🔹 Advanced retrieval techniques including dense retrieval, sparse retrieval, and re-ranking One of the most valuable lessons for me was understanding how to effectively evaluate RAG systems. Learning how to select the right evaluation metrics for different use cases provided a much clearer framework for measuring system performance beyond just subjective response quality. The cohort also reinforced the importance of: ✅ Separating retrieval and generation evaluations ✅ Building deterministic and measurable AI systems ✅ Optimizing context to reduce token costs ✅ Balancing accuracy, latency, and cost in enterprise RAG applications The concepts and evaluation frameworks covered during the cohort are already influencing how I approach the chatbot project I'm currently working on. They have helped me think more systematically about retrieval quality, evaluation strategies, context engineering, latency, and overall system design, enabling me to make more informed architectural decisions and build a more reliable user experience. What I appreciate most about this cohort is that it's not just about consuming content—it's about learning and building in public. Listening to the questions, perspectives, and problem-solving approaches of fellow cohort members has been just as valuable as the sessions themselves. Many times, their questions have challenged my assumptions, reshaped my thinking, and exposed me to perspectives I wouldn't have considered on my own. That collaborative learning environment provides a completely different experience compared to watching recorded lectures. The discussions, debates, and shared experiences make the concepts much more practical and memorable. A huge thank you to Aishwarya Srinivasan, Arvind Narayanamurthy and The Gen Academy team for creating such a thoughtful and engaging learning experience. Looking forward to continuing the journey and applying these learnings to real-world AI applications. As part of the first week's Vibe Coding challenge, I successfully built and submitted my project on time, but never got around to posting it here. Today, I am excited to share it with the community: 🔗 Application: lnkd.in/d2ZxAffR 🎥 Demo: lnkd.in/duuHZ8yR 💻 Codebase: lnkd.in/dHy8Z9pk #AgenticAI #RAG #GenerativeAI #LLM #AIEngineering #EnterpriseAI #LearningInPublic #TheGenAcademy

I have a new 5:00 a.m. boss. Read Build with All 2nd Edition to find out who? Twenty years into my career, after leading enterprise transformations, speaking about AI, and advising leaders on the future of work, I found myself back where it all started: installing an IDE before sunrise. What happened next surprised me. A 20-year-old teaching a 20-year IT professional. I’m becoming a beginner again. And a powerful reminder that the biggest barrier to AI adoption isn't technology; it's behavior. Over the past several weeks, I joined the Mastering Agentic AI cohort from The Gen Academy created by Aishwarya Srinivasan and Arvind Narayanamurthy. Together, they've built more than a learning experience; they've created a community where builders, leaders, students, and practitioners learn side-by-side while exploring what is really happening under the hood of AI. Key topics explored include: ✅ Why every leader should become a beginner again ✅ What a 20-year-old mentor S.Nithish Kumar taught me about leadership ✅ How I built a multi-agent AI solution before work ✅ Why "understanding AI" and "building with AI" are very different things ✅ The Triple AI framework: Actual Intelligence × Artificial Intelligence × All Intelligence If you've ever felt too senior to start over, too busy to learn, or too experienced to ask beginner questions, this story is for you. Subscribe to Build With All - lnkd.in/gkrc_v-p For one practical AI use case, one EQ insight, one stranger's wisdom, and one action you can take this week. Because the future isn't built by the people who watch from a distance; it's built by the people willing to install the IDE again.

Over the last week, I’ve been working through GenAcademy Agentic AI course and spending time building an Agentic RAG application focused on cybersecurity use cases. What stood out to me is how different Agentic RAG feels compared to a traditional RAG implementation. One thing that became clear very quickly is that the value isn’t just in generating answers. It’s in connecting knowledge, tools, and data that already exist across an organization. Thinking back to my deployment days, some of the most challenging troubleshooting sessions happened during production hours. The information I needed was always there, but it was spread across log management platforms, monitoring tools, documentation repositories, ticketing systems, and chat threads. That’s what makes Agentic RAG interesting to me. Instead of hunting through multiple systems, you can ask a question and have an agent gather the relevant information, connect the dots, and return a meaningful answer with context. Here is a snippet of the Known Vulnerability Finder App - collecting information from CISA and other sources that I'm working on. Huge shoutout to Aishwarya Srinivasan and Arvind Narayanamurthy for breaking down the concepts that are easily understandable. Kudos to the TA team at Gen Academy for helping us with our questions during the sessions. Also, all the peers in this session for sharing questions, feedback, ideas. Julia Suzuki Sham Sundar Hassan Chikkegowda Gaurav Munjal Swati Aggarwal

I built a fully local AI teaching assistant for course videos. No paid APIs. No cloud vector database. No data leaving my machine. The problem was simple: when revising from an online course, I often remembered that a concept was explained somewhere across the videos, but not exactly where. So I built a system where I can ask: What is a set in Python? And it returns the exact video, timestamp, matching transcript context, and an AI-generated answer grounded only in that transcript. The pipeline: Converted 13 MP4 course videos into MP3 audio using ffmpeg Transcribed the audio locally with OpenAI Whisper Stored timestamped transcript chunks as JSON Generated local embeddings with Ollama + BGE-M3 Stored 2008 embedded chunks in a Pandas DataFrame Used cosine similarity with NumPy and scikit-learn to retrieve the top matches Added local answer generation with Ollama models like llama3.2 and DeepSeek R1 Wrapped everything in a Streamlit UI The most interesting part was realizing that the AI model was only the last mile. The real work was engineering the pipeline: making the transcripts timestamp-aware choosing the right chunk size merging short Whisper segments into richer context keeping retrieval simple and explainable writing prompts that force the LLM to stay grounded in the provided video context This project proved something important to me: You do not always need OpenAI APIs, LangChain, Pinecone, or cloud infrastructure to build a useful RAG system. With open-source tools and some engineering discipline, a complete AI workflow can run privately on your own machine. Stack used -> - ffmpeg - OpenAI Whisper running locally - Ollama - BGE-M3 embeddings - llama3.2 / DeepSeek R1 - Pandas - NumPy - scikit-learn - Streamlit Small project, but a big learning curve. A special thanks to Aishwarya Srinivasan and Arvind Narayanamurthy for the motivation throughout this cohort. This cohort pushed me to move faster than I normally would. Honestly, I probably would have stretched this project over 30 days, kept overthinking the architecture, and refined it endlessly. But because we were already learning and working through RAG concepts in Week 2, I decided to build it properly and ship it early. What could have taken a month ended up coming together in just 3-4 days. That momentum made a real difference. Thank you both for creating the kind of learning environment that pushes people to stop just thinking about ideas and actually build them. The GitHub link to the project is in the comments. #AI #RAG #MachineLearning #Python #Ollama #Whisper #Streamlit #OpenSourceAI #LocalAI #DataPrivacy #buildInPublic

RAG-Powered FIFA World Cup Q&A System Evaluation I built a RAG-powered FIFA World Cup Q&A system and ran it through proper evaluation. lnkd.in/gZnFk924 Most RAG tutorials stop at "it works." I wanted to know how well — so I built a golden dataset, wired up RAGAS evaluation with Claude as judge, and ran 17 test cases against it. What's under the hood — Hybrid Retrieval: dual Pinecone indexes (dense + BM25 sparse), alpha- weighted score fusion, refusal threshold when evidence is weak — LangGraph pipeline: typed StateGraph with query rewriting, out-of-scope detection, and conditional routing — Claude: answer generation with strict citation grounding — no hallucination without a source — corpus: 1,466 documents across 22 World Cup editions (1930–2022) What I actually learned building this 💡 Hybrid retrieval isn't optional for factual QA — BM25 catches exact matches (player names, years, scores) that dense embeddings miss 💡 Query rewriting changes everything — "Who won the 19th World Cup?" fails without mapping ordinals to years first 💡 A refusal threshold is more important than retrieval quality — knowing when not to answer is what separates a demo from a reliable system 💡 RAGAS Answer Relevancy penalizes you when your answer doesn't echo the question's key terms — even if it's factually correct. That one insight changed how I write generation prompts 💡 Evaluation-driven development works — writing the golden dataset before tuning forced me to be honest about what the system could and couldn't do. Thanks to Aishwarya Srinivasan and Arvind Narayanamurthy for sharing their knowledge and insights through the amazing cohort. Always happy to connect and learn with people building AI, RAG, and agentic systems. #RAG #AgenticAI LangChain LangGraph AI Pinecone Anthropic #RAGAS #GenerativeAI #FIFAWorldCup

Excited to share that I’ve completed the Mastering Agentic AI Certification by The Gen Academy! 🎉 This was far from a conventional certification. The program kept us genuinely engaged throughout — weekly live sessions packed with deep, practical content, paired with hands-on projects every week that pushed us to actually build, not just learn theory. It's been a fantastic learning experience, and I'm looking forward to applying everything I've picked up to build real-world AI and agentic solutions in my career. Grateful and thank you so much Aishwarya Srinivasan, Arvind Narayanamurthy and The Gen Academy team for putting together such a valuable cohort #AgenticAI #ArtificialIntelligence #GenerativeAI #MachineLearning #LLM #RAG #MultiAgentSystems #AIDataEngineer #DataEngineering #OpenToWork #Hiring #TechJobs #DataEngineerJobs #AIJobs

🚀 Excited to share that I have successfully completed the first week of my Agentic AI Certification journey! A huge shout-out to The Gen Academy team, Aishwarya Srinivasan and Arvind Narayanamurthy for putting together such high-quality content. The sessions have been incredibly insightful, practical, and packed with valuable learnings. What makes this journey even more special is that I've been following this community since their Context Engineering workshop. That workshop played a significant role in kickstarting my AI journey and gave me the confidence to move from learning concepts to actually building solutions. Over time, the knowledge and guidance I've gained have helped me develop AI-powered prototypes and proof-of-concepts that I've been able to showcase at work. Seeing ideas turn into working solutions has been both rewarding and motivating. The first week of the certification has reinforced why I chose to continue learning with this community. I'm looking forward to the upcoming sessions, hands-on projects, deeper technical discussions, and the opportunity to collaborate with fellow learners. Excited for what's ahead and grateful to be part of this amazing community! 🙌The Gen Academy #AgenticAI #ArtificialIntelligence #AIEngineering #GenerativeAI #LLM #MachineLearning #LearningJourney #TechCommunity #Innovation #ContinuousLearning

My latest deep dive in the Mastering Agentic AI Bootcamp took my understanding of one of the most powerful tools in modern AI to the next level: 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴 🚀 This week pushed me to think more critically about when and why to use fine-tuning versus other approaches. One of the most clarifying moments was developing a stronger intuition around the decision of when to go with 𝗽𝗿𝗼𝗺𝗽𝘁𝗶𝗻𝗴 💬, when 𝗥𝗔𝗚 🔍 is the right call, and when 𝗳𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 ⚙️ is genuinely the answer. Exploring the different fine-tuning methods was equally eye-opening. Understanding the distinctions between 𝗦𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗲𝗱 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴, 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴, and 𝗣𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿-𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴 gave me a much clearer picture of the tradeoffs involved and the scenarios where each approach truly shines. To put the learning into practice, we built an 𝗜𝗧 𝗛𝗲𝗹𝗽𝗱𝗲𝘀𝗸 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴 𝗣𝗶𝗽𝗲𝗹𝗶𝗻𝗲 using 𝗦𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗲𝗱 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗟𝗼𝗥𝗔 on the 𝗤𝘄𝗲𝗻𝟯 𝟭.𝟳𝗕 base model. The image below walks through the full pipeline and the numbers speak for themselves 📈. A huge shout-out to Aishwarya Srinivasan and Arvind Narayanamurthy for the clarity they brought to every session 👏. Their ability to make complex concepts click and connect theory to real implementation made all the difference. This is exactly what makes the The Gen Academy bootcamp so valuable. More to come as we head into the final stretch! 🎯 #AgenticAI #FineTuning #LLM #MachineLearning #AIEngineering #MasteringAgenticAI #GenAcademy #AIBootcamp #ContinuousLearning

Agentic AI is easy to demo. Getting it to actually hold up in the real world is the hard part. I just wrapped up the inaugural May cohort of the Mastering Agentic AI certification with Aishwarya, Arvind and the team at The Gen Academy. I’m genuinely grateful for the practical, real-world experience they brought into every session. Not just theory but actual lessons from systems they’ve built, tested, and broken. I picked the live cohort on purpose. I wanted to learn from practitioners actively building these systems, rather than rely on pre-recorded content. That meant real questions, real debugging, and honest conversations about the tradeoffs involved. Their live sessions were engaging, fast-paced, and packed with practical frameworks I could apply immediately. My biggest takeaway was this. Connecting an LLM to a few tools can create an impressive demo. It doesn’t create a dependable system. The real work, the part that determines whether a system survives contact with actual users, lies in architecture, context engineering, evaluation, observability, security, and governance. We explored everything from LLM fundamentals, prompting, function calling, and structured outputs to RAG, MCP, A2A, LangChain, and LangGraph. We also worked with fine-tuning, LoRA/QLoRA, and local models. Just as importantly, we spent meaningful time on what often gets overlooked, like golden datasets, failure analysis, monitoring, LLM-as-a-judge, prompt injection, privacy, guardrails, and red teaming. Nothing stayed theoretical. Through hands-on builds, an end-to-end capstone, and guest sessions and support from teams across the AI ecosystem (NVIDIA, Pinecone, OpenAI, LangChain, Replit, Fireworks AI, and many others), we built real systems, watched them fail, diagnosed why, and improved them. That discipline matters for businesses. Air Canada learned this when its chatbot gave a customer inaccurate information about claiming a bereavement fare after traveling. The customer relied on it, and a tribunal ultimately held the airline responsible. One incorrect response created real financial and reputational consequences. AI only creates meaningful value when it's reliable, secure, measurable, and actually ready for real customers and real decisions. If you want to move beyond being AI-curious and learn to build systems you would confidently stand behind, I highly recommend this program. Looking forward to putting these lessons into practice. Course link in the comments. #AgenticAI #AIEngineering #GenerativeAI #AIStrategy #ContinuousLearning

lnkd.in/ePEBSk_5 Excited to share that I've completed Mastering Agentic AI from The Gen Academy! 🎓 This was so much more than a course. I didn't just learn concepts — I got to build real projects, push my thinking, and grow alongside an incredible group of people. A heartfelt thank you to Aishwarya Srinivasan and Arvind Narayanamurthy for your guidance, patience, and for making complex ideas click. Your mentorship shaped how I approach agentic AI. To the interns — thank you for your support and for always being ready to help. And to my fellow cohort members — you inspired me every single step of the way. Learning with you, building with you, and being challenged by your ideas made this journey unforgettable. Grateful for everyone who made this experience what it was. Onward to building more! 🚀 #AgenticAI #TheGenAcademy #AI #ContinuousLearning #Grateful

When Aishwarya Srinivasan announced the first cohort of Mastering Agentic AI, I knew I wanted to be part of it. I wanted to move beyond learning AI concepts and start building real AI applications. Throughout the program, we didn't just learn AI concepts—we built hands-on AI projects, experimented with multi-agent workflows, explored RAG, evaluations, observability, fine-tuning, and AI security. Today, I'm happy to have completed the program. A heartfelt thank you to Aishwarya Srinivasan, Arvind Narayanamurthy, and The Gen Academy for creating a hands-on learning experience that transformed complex AI concepts into practical skills through real projects. Looking forward to continuing to build, learn, and apply these skills to solve real-world problems. #AgenticAI #GenerativeAI #LLMs #RAG #AIEvaluation #ContinuousLearning #LearningByBuilding

I’ve completed the Mastering Agentic AI masterclass with The Gen Academy. After years of working with Java and microservices, I’ve been steadily moving deeper into RAG and agentic AI. This masterclass helped me think more clearly about what it takes to move beyond an LLM demo and build agents that can reason, use tools, and produce reliable results. The certificate is a nice milestone, but I’m more interested in what comes next: applying these ideas, building useful systems, and learning where agents genuinely add value. Thank you Aishwarya Srinivasan and Arvind Narayanamurthy for putting the program together, and to everyone in the cohort who made the learning experience better. Back to building. #AgenticAI #GenerativeAI #AIEngineering

🎓 A meaningful milestone in my AI learning journey Happy to share that I’ve officially completed the Mastering Agentic AI cohort and received two certificates: One from Maven and another from The Gen Academy. The Gen Academy certificate feels especially meaningful because it represents the hands-on work that went into the program. Throughout the program, the coursework, assignments, and hands-on projects pushed me to apply what I was learning rather than just attend the sessions. Really grateful to Aishwarya Srinivasan, Arvind Narayanamurthy, and The Gen Academy team for making the first cohort such a great learning experience. Definitely not done learning yet. 😊 Still learning. Still experimenting. Still building. 🚀 #AgenticAI #GenerativeAI #AIAgents #AIEngineering #GenAI #ContinuousLearning #SoftwareEngineering

🎓 Excited to share that I've completed the 6-week hands-on "Mastering Agentic AI" masterclass by The Gen Academy! This wasn't a course where you just read about AI agents — you actually build them. Over six weeks, I got to: 🔹 Understand how AI agents really work under the hood 🔹 Build working agents, not just study the theory 🔹 Dive into model evaluation 🔹 Explore the various tools, frameworks, and techniques available in the market today 🔹 Learn alongside an incredible cohort — seeing how industry peers approach ideation and problem-solving was a huge bonus A big thank you to Aishwarya Srinivasan, Arvind Narayanamurthy, and the entire The Gen Academy team. The coursework was thorough and well-structured, and the whole experience pushes you to get comfortable by getting uncomfortable — which is exactly where the real learning happens. #AgenticAI #AI #ModelEvaluation #ArtificialIntelligence #ContinuousLearning #TheGenAcademy

Is your LinkedIn feed also flooded with Agentic AI posts? Mine certainly is! Every day, I see articles, demos, architecture diagrams, GitHub repositories, and success stories about Agentic AI. It's clear that we're at a pivotal moment where intelligent agents are redefining how we build software and solve business problems. Curious to learn beyond the hype, I enrolled in The Gen Academy's Mastering Agentic AI course. I initially expected a recap of what I'd already learned from YouTube videos, blogs, and research papers—but the experience exceeded my expectations. The first session by Aishwarya stood out by focusing on the Agentic AI technology stack and, more importantly, the architectural principles behind it. It emphasized that while frameworks evolve rapidly, the fundamentals—reasoning, planning, memory, retrieval, orchestration, tool integration, evaluation, and governance—are what truly matter. As a Software Engineer, I found this perspective invaluable. Agentic AI is much more than prompt engineering. Building reliable AI systems requires a strong foundation in software architecture, APIs, RAG, vector databases, observability, security, and scalable engineering practices. For technology leaders, the opportunity extends beyond technology. The real value lies in using AI to automate workflows, improve productivity, enhance customer experiences, and create measurable business outcomes. Success won't come from chasing every new framework—it will come from combining strong engineering fundamentals with responsible AI adoption. Having started my career during the dot-com era, I see AI as another transformational shift—one that is augmenting not only software but also human creativity and decision-making. Whether you're a developer, architect, engineering manager, or technology leader, now is the time to invest in understanding the principles behind Agentic AI. Learn continuously, experiment with real-world use cases, and focus on creating meaningful business value. A very special thank you to The Gen Academy and team for delivering a well-structured Mastering Agentic AI program that bridges technical depth with practical application. The future belongs to engineers and leaders who understand AI and work together to build the next generation of intelligent systems. #AgenticAI #ArtificialIntelligence #GenerativeAI #SoftwareEngineering #TechnologyLeadership #AIArchitecture #RAG #Innovation #ContinuousLearning #TheGenAcademy

Curiosity has always been one of my biggest motivations for learning. I’m excited to share that I’ve completed the Mastering Agentic AI certification from The Gen Academy . A big thank you to Aishwarya Srinivasan , Arvind Narayanamurthy , and the team for creating such an engaging learning experience. One of my biggest takeaways was learning to think beyond individual models and focus on designing AI systems: combining agents, retrieval, tools, and evaluation into reliable workflows. I’m excited to bring these concepts into future AI projects and continue exploring how Agentic AI complements traditional machine learning. What Agentic AI application do you think will have the biggest real-world impact over the next few years? #AgenticAI

I initially came across Aishwarya Srinivasan content while trying to familiarize myself with AI terminology and better understand the rapidly evolving AI landscape. What stood out immediately was her ability to break down complex concepts into digestible, practical learning. The Mastering Agentic AI course by Aishwarya Srinivasan and Arvind Narayanamurthy has been incredibly well structured — not just introducing AI jargon, but providing the context needed to truly understand and apply concepts such as tokens, embeddings, orchestration frameworks like LangChain, and the importance of selecting the right models and tools for the right use cases. What I particularly appreciate is how the course introduces the evolution of Large Language Models (LLMs), helping learners understand not only where AI stands today, but also how we arrived here and where this technology may be heading. That broader perspective makes the learning more meaningful and helps connect foundational concepts with the future of agentic systems. Concepts like neural networks and embeddings can often feel abstract when encountered without context. The course made these ideas approachable and actionable, helping bridge the gap between theory and practical application and giving me greater confidence to explore AI-powered systems more deeply. One of our first assignments was to vibe code an application. I was able to get a functional proof of concept working from a single well-structured prompt. The experience was both exciting and eye-opening. Today’s models have matured significantly,when provided with clear requirements, sufficient context, and the right constraints, they can make meaningful implementation decisions and accelerate development dramatically. As a full-stack engineer with 15 years of experience building production systems, what stood out to me was not just the speed of prototyping, but how understanding model capabilities, tokens, and orchestration frameworks can meaningfully accelerate early-stage development. At the same time, this experience reinforced an important distinction: AI-generated POCs are not the same as enterprise-grade systems. Building production software still requires architectural thinking, domain expertise, integrations, governance, collaboration, and countless conversations across teams to ensure systems are secure, scalable, and reliable. Just as frameworks like React or Angular transformed how we build interactive applications and cloud technologies simplified infrastructure and networking, AI represents another architectural shift one that helps us rethink workflows, augment engineering capabilities, and move toward greater automation. AI is not replacing engineering expertise — it is amplifying it. Excited to continue learning, experimenting, and building in this space. Can’t wait to learn so much more in the next week weeks. #AgenticAI #LLM #AIEngineering #AIAgents #TechLearning #AIForDevelopers #ContinuousLearning #MasteringAgenticAI

Excited to share that I’ve completed the Mastering Agentic AI certification! Looking forward to applying these skills to build smarter AI products and agentic systems. 🚀 Many thanks to Aishwarya Srinivasan, Arvind Narayanamurthy and The Gen Academy team!

Building an AI agent is only the first step. Proving that it works is where the real engineering begins. For my latest project, I extended TrustLayer AI, my multi-agent Python code review system, by focusing on evaluation and observability rather than adding new features. To answer the question, "How do I know my AI agent is actually reliable?", I: → Built a 40-case golden dataset covering happy paths, edge cases, known failures, and adversarial scenarios. → Defined measurable evaluation metrics instead of relying on "it looks good." → Designed an observability workflow inspired by LangSmith concepts to trace agent execution, workflow decisions, latency, and evaluation results. → Ran baseline evaluations, analyzed failure patterns, implemented targeted improvements, and measured the impact against the same benchmark. This project reinforced an important idea: "Building AI is exciting. Measuring, understanding, and continuously improving AI is engineering." Thank you Aishwarya Srinivasan and Arvind Narayanamurthy for another practical project that challenged us to think beyond building AI demos and toward building more reliable AI systems. I'm looking forward to applying these evaluation and observability practices as I continue building more reliable AI systems. #AI #AIEngineering #AgenticAI #GenerativeAI #AIEvals #LangSmith #Observability #Python #ContinuousLearning

Grateful to have completed the Mastering Agentic AI masterclass with The Gen Academy. Thank you Aishwarya Srinivasan Arvind Narayanamurthy. Excited to keep learning and building 🚀.

🎓 Learning milestone accomplished — completed the Mastering Agentic AI certification by The Gen Academy! Coming into Mastering Agentic AI, I was grounded in the basic concepts of machine learning through my previous learning as part of the PG Program in AI & ML from UT Austin. I understood how models are built and how they work under the hood — supervised and unsupervised learning, deep learning, NLP, the core ideas behind modern AI. But there was a gap. Understanding how models work is one thing; engineering a production-grade RAG and agentic systems around them is another entirely — and this course was the fast, structured path to bridging it. The LangChain and LangGraph deep dive gave me the tools to architect stateful, multi-agent systems with human-in-the-loop control. Then it went where most courses stop: evaluation with golden datasets and LLM-as-a-judge, observability and monitoring to keep systems reliable in production, and eval loops that catch regressions and hallucinations before they reach users — plus guardrails and prompt injection handling that turned "I built a demo" into "I can run this safely." I went from understanding AI concepts to shipping an end-to-end production-grade agentic application. And because every session is live, the real learning happened in the moments where I could ask questions and build alongside. Thank you Aishwarya Srinivasan and Arvind Narayanamurthy for showing how to go from reading blogs and following tutorials to actually building the real thing. 🙏

One of the projects I was fortunate to work on during the The Gen Academy cohort was RouteIQ—a solution designed to help travelers plan their entire trip without having to open 30 different tabs. My role focused on identifying the market opportunity, understanding customer pain points, and defining the user journey to ensure we were solving a meaningful problem. The team then brought that vision to life through a GraphRAG-powered day planner with specialized agents for routing, nearby recommendations, and restaurant suggestions, supported by a thoughtfully designed evaluation framework. This project would not have been possible without the guidance and support of Aishwarya Srinivasan and Arvind Narayanamurthy, whose mentorship shaped our thinking throughout the journey. What stood out most was their belief that anyone can learn AI, understand the architecture, and build with it regardless of their background. As someone who started this cohort feeling overwhelmed, I can confidently say they stayed true to that promise. I'm also incredibly grateful to Guruprasad Balasubramaniam for his technical architecture leadership and to Siddardha Polineni and Anna Ayvazyan for their valuable feedback in refining the architecture and strengthening our evaluation approach. This experience was a perfect example of how product thinking, AI innovation, technical excellence, and strong mentorship can come together to build something impactful. But this is only the beginning. Getting the intent right was the first half of the challenge. Getting the trip to feel right for that intent is the next half—and that's exactly where our next iteration is focused. We're already exploring ways to make RouteIQ even more personalized, context-aware, and intuitive. Stay tuned!!

Seven weeks ago, I could build an agent. I couldn't tell you whether it should be trusted. That's the gap this bootcamp actually closed. The Gen Academy's Mastering Agentic AI Bootcamp runs in two halves. The first half is construction: RAG that actually retrieves: hybrid search, reranking, graph retrieval, then agents. ReAct, plan-and-execute, MCP, LangGraph, multi-agent orchestration, and the cost and latency tradeoffs that decide whether any of it survives real traffic. The second half is everything that decides whether you ship it. Evals and LLM-as-judge. Tracing and regression pipelines. Fine-tuning, and the harder question of when it's actually the right call. Then security: prompt injection, guardrails, data privacy, governance. Three things from that second half changed how I work. I write the negative controls now, the cases built to make my own system look bad, before anyone else finds them. I read the failures by hand instead of trusting the eval score, because I've watched a judge score confidently and get it wrong. And I stopped treating a guardrail as authorization. If the model is the only thing standing between a user and someone else's data, the system is already broken. No prompt fixes that. Group demos Saturday. Last day yesterday. Thank you, Aishwarya Srinivasan and Arvind Narayanamurthy! You spent half the curriculum teaching us to build these systems and the other half teaching us not to take them at their word. That second half is the part that shows up in the job. Verde Bowl, my security capstone, gets its own post next week. #AgenticAI #AIEngineering #LLM #AISecurity

🎓 Week 2 of Mastering Agentic AI Bootcamp - Shipped! ✨ 𝗙𝗼𝗰𝘂𝘀: Context Engineering, Agentic RAG & RAG Evals. 𝗣𝗵𝗶𝗹𝗼𝘀𝗼𝗽𝗵𝘆: The best way to learn is to build, so here it is #buildinpublic 𝗠𝗲𝗲𝘁 𝗠𝗜𝗥𝗔 - 𝗠𝗲𝗿𝗶𝗱𝗶𝗮𝗻 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝘁 𝗥𝗼𝘂𝘁𝗲 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 ✈️ A production-grade Agentic RAG AI Assistant designed for a fictional airline (Meridian Airlines). Every answer is sourced from 56 official airline documents. No hallucinations. No training data leakage. No guesswork. Just verified facts, cited inline, every single time. 📚 𝗪𝗵𝗮𝘁 𝗠𝗮𝗸𝗲𝘀 𝗠𝗜𝗥𝗔 𝗗𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁? Most RAG systems follow a simple pattern: 𝘦𝘮𝘣𝘦𝘥 → 𝘳𝘦𝘵𝘳𝘪𝘦𝘷𝘦 → 𝘢𝘯𝘴𝘸𝘦𝘳. MIRA flips the script with intelligent reasoning at every layer: 🗺️ 𝗞𝗻𝗼𝘄𝘀 𝗪𝗵𝗮𝘁 𝘁𝗼 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗲 Before fetching a single document, MIRA selects the most relevant tools from a library of 16 topic-scoped retrievers. Baggage questions go to the baggage tool. Loyalty program questions to the FFP tool. Zero wasted retrieval calls. ⚡ 𝗥𝘂𝗻𝘀 𝗦𝗺𝗮𝗿𝘁, 𝗡𝗼𝘁 𝗝𝘂𝘀𝘁 𝗙𝗮𝘀𝘁 BM25 (lexical search) and vector search execute in parallel. ID-based queries? BM25 finds exact matches instantly. Semantic questions? Vector search handles nuance. Both run simultaneously—the best of hybrid search without the overhead. 🎯 𝗖𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝘁 𝗘𝗻𝗼𝘂𝗴𝗵 𝘁𝗼 𝗦𝗸𝗶𝗽 𝗪𝗼𝗿𝗸 Re-ranking is expensive (~900ms per query). MIRA asks: “Am I already confident?” If the top result scores ≥0.65 cosine similarity, re-ranking is skipped entirely. Intelligence isn't about doing everything—it's about doing what matters. 💬 𝗢𝗻𝗲 𝗖𝗮𝗹𝗹. 𝗢𝗻𝗲 𝗔𝗻𝘀𝘄𝗲𝗿. 𝗙𝘂𝗹𝗹 𝗖𝗼𝗻𝘁𝗲𝘅𝘁. All retrieved context flows into a single LLM call. No chain-of-thought loops. No multi-turn reasoning overhead. Synthesized answer with citations. Done. 𝗘𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲 𝗜𝘁 𝗬𝗼𝘂𝗿𝘀𝗲𝗹𝗳 🖥️ Live App: lnkd.in/gdUw7c-F (best viewed on laptop/desktop) 💻 Github: lnkd.in/g7GcExEW Huge thanks to Aishwarya Srinivasan and Arvind Narayanamurthy for this. Every session has been deeply technical, genuinely engaging, and impossible not to build after. The structured depth combined with the #buildinpublic philosophy is exactly what turns learning into shipped projects! #BuildInPublic #AgenticAI #RAG #ContextEngineering #LangGraph #Pinecone #NVIDIA #Streamlit #GenerativeAI #AIEngineering #MasteringAgenticAI

Strong learners enrolled yesterday for the July 5 Cohort of the Agentic AI Enterprise Mastery Bootcamp. And I am genuinely grateful for the kind of learners this cohort is attracting. People who are not just looking for another AI tutorial. People who want to understand how agents, tools, RAG, memory, APIs, guardrails, observability, and deployment readiness come together into one production-style system. That is the direction of this bootcamp. Not random demos. Not surface-level tool usage. Not “build this in 10 minutes” learning. The goal is to build one production-style Agentic AI system end to end — layer by layer — and develop the judgment to explain, debug, and improve it. This coming week, I’ll also be sharing more learning resources on our YouTube channel around Agentic AI systems, reliable tool design, RAG as grounding, observability, and production readiness. If this direction aligns with where you want to grow, you can check out the July 5 Cohort details here: lnkd.in/gYUybcZk I also recorded a walkthrough of what you actually get inside the bootcamp: lnkd.in/gebgtexg Systems ship. Demos don’t.

🎓 Excited to share that I've completed the Mastering Agentic AI Bootcamp! As a full-stack engineer, I've spent years building software systems. This year, I wanted to go beyond simply using AI tools and understand how production AI systems are designed, evaluated, and deployed. 𝗪𝗵𝘆 𝗜 𝗲𝗻𝗿𝗼𝗹𝗹𝗲𝗱 I wanted to move beyond using LLMs and learn how to engineer production AI systems—from Agentic AI and RAG to memory, evaluations, and the engineering trade-offs behind building reliable, scalable applications across quality, latency, cost, and security. 𝗪𝗵𝗮𝘁 𝗜 𝗹𝗲𝗮𝗿𝗻𝗲𝗱 Over the past few weeks, I explored and built with: → Agentic AI architectures & multi-agent orchestration → Retrieval-Augmented Generation (RAG), hybrid search, metadata filtering, reranking & evaluation strategies → AI system design & production architecture → Memory, tool calling, MCP & reasoning workflows → Embeddings, vector databases & retrieval optimization → Prompt engineering, AI observability, evaluations & golden datasets → AI security, guardrails, governance & production readiness → Fine-tuning strategies, local models & choosing the right model → LLM fundamentals, inference & the trade-offs behind modern AI systems What stood out most was that we weren't given blueprints to copy. Instead, we were encouraged to design our own architectures, make engineering trade-offs, experiment, iterate, and understand why one approach works better than another. That mindset made all the difference. The bootcamp also provided access to production-grade tools, guest lectures, curated learning resources, and an incredibly supportive community. 𝗪𝗵𝗮𝘁 𝗜'𝗹𝗹 𝗮𝗽𝗽𝗹𝘆 The biggest takeaway for me is that building AI applications isn't just about connecting an LLM to an API. It's about designing reliable systems with the right retrieval strategy, memory, orchestration, evaluation, observability, and security. I'm already applying these concepts by building: ✅ Agentic RAG systems ✅ Evaluation pipelines with golden datasets ✅ Multi-agent workflows More importantly, I'm bringing this architectural mindset into my day-to-day engineering work identifying where AI genuinely creates value and where simpler solutions are the better choice. As AI continues to evolve, I want to keep building systems that are not only intelligent, but also scalable, reliable, and production-ready. A huge thank you to Aishwarya Srinivasan and Arvind Narayanamurthy for creating a program that combines strong engineering fundamentals with practical, hands-on learning. Excited to keep learning, experimenting, and building. 🚀 #AgenticAI #AIEngineering #AISecurity #ProductionAI #SystemDesign #RAG #MultiAgentSystems #LLMs #AIArchitecture #GenAI #BuildInPublic #ContinuousLearning lnkd.in/gT-y9Vem

Last week’s session in the Mastering Agentic AI journey was another power-packed one! 🚀 A huge appreciation to aishwarya and Arvind Narayanamurthy for making complex concepts easy to understand through clear explanations and practical examples. The session covered different RAG (Retrieval-Augmented Generation) mechanisms, including: 🔹 Traditional RAG 🔹 Agentic RAG 🔹 Graph RAG 🔹 RAG evaluation concepts 🔹 Introduction to LangGraph and how it can be used to build more controlled, stateful agentic workflows What I especially appreciated was that the session wasn’t just about what these concepts are, but also why and when we would use each approach. Connecting the architecture with real-world examples made the differences much easier to understand. Thank you, Aishwarya and Arvind, for another session packed with knowledge — a true protein-packed learning session for the AI brain! 🧠💪 Looking forward to applying these concepts at work. #AgenticAI #GenerativeAI #RAG #AgenticRAG #GraphRAG #LangGraph #AIEngineering #LLM #LearningJourney #ArtificialIntelligence

Just wrapped a genuinely useful session in the Agentic AI Bootcamp: a masterclass on "Agentic AI System Design" by Aishwarya Srinivasan, Arvind Narayanamurthy, and Arpit Chaudhary. Grateful for how openly they shared the hard-won stuff. What set it apart: instead of more prompt tips, they reframed the whole problem. "Can the model do it?" Usually yes. The real engineering is the system around it. The LLM is slow, costly, occasionally wrong, and stateless, so your job is to wrap it in deterministic, tested, monitored code and make the system reliable, cost-effective, fast, and context-aware. A few ideas I'm taking with me: - One big prompt becomes a pipeline of small, schema-contracted steps you can test and retry individually. - Right model per step. Spend only on the reasoning that changes the outcome. - Memory is your job, not the model's: four access patterns, four stores. - Everything that touches the model is attacker input. The LLM suggests, the code validates, and the user authorizes. The best part was walking through Eikos, a real shipped WhatsApp appointment-booking product, end-to-end on AWS for a near-zero infra bill at launch. Seeing the actual architecture, not a toy demo, is what made it click. Thank you both for teaching this so generously. Highly recommend it to anyone serious about building, not just prototyping. #AgenticAI #AIEngineering #LLM #SystemDesign #BuildInPublic

I spent two weeks debugging a RAG system where retrieval scores looked good on paper but the answers were consistently mediocre. The chunks were relevant. The model was capable. I could not figure out what was wrong until I opened a few chunks and read them. Half started mid-sentence because they were sliced at a fixed character count. Others contained a heading from one section and body text from the next. The system was retrieving the right topics but feeding the model mangled fragments that no human could make sense of either. That experience taught me something I now treat as a rule: retrieval quality is capped by chunking quality. You can have the best embedding model and the smartest reranker, but if your chunks are incoherent, everything downstream suffers. Here is the core problem: Documents have structure i.e. Paragraphs, sections, headings, lists, tables. That structure carries meaning. When you slice a document into chunks for embedding, you are deciding where meaning begins and ends. Get it right and each chunk is a self-contained thought that embeds well and retrieves well. Get it wrong and you have fragments too small to mean anything, or too large to be specific, or split across ideas in ways that confuse both embedding and generator. The naive approach is fixed-size chunking. Take every 512 tokens regardless of content. Fast, easy to implement, and terrible for retrieval quality. Sentence boundaries break. A definition gets separated from its explanation. A question splits from its answer. The embedding represents a garbled fragment, and similarity search dutifully returns the closest garble. Better approaches chunk by structure. Split at paragraph boundaries, section headers, or natural topic shifts. Overlap adjacent chunks by 10 to 20 percent so ideas spanning a boundary are captured on both sides. Prepend headings to each chunk within their section, so a chunk about "return policy exceptions" always carries the heading "Return Policy" with it for context. The hardest part is that optimal chunking depends on your specific documents and queries. A legal brief with long paragraphs needs different sizes than a FAQ with short answers. API docs with code blocks need structure-aware splitting that keeps code with its description. No universal right answer exists, which means you test and iterate. The metric that matters is retrieval precision: given a query, does the top chunk actually contain the answer in a readable, self-contained form. If you only measure end-to-end answer quality, you cannot tell whether a bad answer was the model's fault or the chunk's fault. Measuring retrieval independently lets you fix the right layer. What chunking strategy are you using, and have you tested whether different sizes change retrieval quality? #ArtificialIntelligence #MachineLearning #LLM #GenerativeAI #AIMLPlaybook

I have been thinking about learning AI for a long time. Finally, I got the opportunity to join the Gen Academy August Cohort. Here is my Week 2 project: APS School Chatbot. The APS Online Tamil School Assistant Bot is designed to answer both general and specific questions from parents. The bot retrieves information from FAQs, service requests, and manuals to provide accurate responses to parents’ questions. It also displays supporting evidence for its answers. If the bot cannot confidently answer a question, it escalates the request to a human administrator instead of providing potentially inaccurate or hallucinated information. The solution architecture for the APS Online Tamil School Assistant Bot consists of four key steps: Semantic Search: Uses ChromaDB and local Hugging Face embeddings to identify records that are semantically similar to the user’s question. BM25 Search: Performs keyword-based matching across documents loaded from ChromaDB. Reciprocal Rank Fusion (RRF): Combines the results from semantic search and BM25 search to produce a unified ranking of the most relevant records. Cosine Similarity Scoring: Evaluates the relevance of the best-matched record. If the similarity score meets the configured confidence threshold, the bot returns the answer along with supporting evidence. Otherwise, the question is escalated to a human administrator for review. Thank you, Aishwarya Srinivasan Arvind Narayanamurthy #RAG #GenerativeAI #AIEngineering #AgenticAI #TheGenAcademy

Week#4 Project - Built an AI evaluation pipeline for e-commerce customer support routing. The system uses LangGraph and GPT-4o-mini to classify support tickets across five domains, with Pydantic for structured outputs. I evaluated 100 labeled tickets using pandas and scikit-learn, then added OpenTelemetry and LangSmith for prediction-level tracing and failure analysis. A focused prompt improvement increased accuracy from 91% to 99%, reducing errors from 9 to 1 with zero regressions. Thanks to Aishwarya Srinivasan Arvind Narayanamurthy and The Gen Academy for their support through "Mastering Agentic AI: Certification by The Gen Academy" course.

I'm four weeks into TheGenAcademy's Mastering Agentic AI course (a couple of sessions still to go) - and it has genuinely been like drinking from a fire hose. Vibe-coding, building RAG pipelines, working with deep agents, running evaluations... in a month, I've gone from reading about these systems to actually building them. Here's why I signed up: I've spent the last 3 years building and running an AI Governance program. But somewhere along the way, I realized I was governing systems I understood mostly from the outside - from policy documents, and risk frameworks, not from the architecture itself. I wanted to close that gap. I needed to understand: -How these systems are actually built, end to end -What goes into the architecture and the decisions behind it -Why evaluations matter — and what "good" evaluation actually looks like in practice -Where the real security and privacy risks live, not just the ones on a checklist The course requires a project submission every week based on that week's material, and I've turned in 4 so far. Each one has been its own crash course - there's a real difference between reviewing an AI system's documentation and having actually built the thing yourself. You start asking sharper questions, spotting gaps that don't show up in a risk matrix, and telling a genuine control apart from a paper one. It's already reframing how I think about scaling the governance program I've built. Responsible AI and AI Governance gave me the "what" and the "why" - this course is giving me the "how," and that combination is exactly what governance functions need more of as these systems get more agentic and autonomous. Over the next few posts, I'll walk through each of the 4 projects - what they were, what they taught me, and where I saw it connect back to governance. Starting with week one, next up. #TheGenAcademy #MasteringAgenticAI A big shoutout to Aishwarya Srinivasan and Arvind Narayanamurthy for their powerhouse course, and keeping things simple

Kubernetes incident history can be valuable. But only if engineers can find the right record when they need it. I built the Platform Reliability Assistant, a RAG application using n8n. It helps engineers search ServiceNow-style Kubernetes incident, problem, and change records using natural language. What it does: • Retrieves records based on symptoms and operational context • Finds previous causes, resolutions, and corrective changes • Shows relationships only when the user asks for them • Handles follow-up questions using conversation memory • Provides record references and Kubernetes documentation links I created a fully synthetic ITSM dataset based on official Kubernetes documentation. No company, customer, or production data was used. How it works: JSONL corpus → validation → chunking → embeddings → hybrid retrieval → evidence selection → cited answer The application uses the Qwen3-Embedding-8B model through Nebius for embedding, Pinecone for vector search, BM25 for keyword matching, and Mem0 for follow-up context and OpenAI gpt-oss-120b through Nebius token factory for generation. My biggest learning was that the model is only one part of a RAG application. Corpus quality, chunking, metadata, retrieval, ranking, and evidence selection have a direct impact on the reliability of the final answer. Evaluating showed me the importance of checking whether the correct record reached the model, instead of treating every incorrect answer as a prompting problem. This was my Week 2 project for the Mastering Agentic AI program. Thank you, Aishwarya Srinivasan, Arvind Narayanamurthy and The Gen Academy for creating an environment where we learn by building, testing, and improving. #RAG #Kubernetes #AIEngineering #ITSM #AgenticAI #ServiceNow #TheGenAcademy

🚀 Turning GenAI Learning into Practical Enterprise Solutions I’m excited to share a milestone in my ongoing Mastering Agentic AI learning journey with The Gen Academy. As part of the program, I recently designed and developed an AI-Enabled Application Management Services (AMS) Incident Analyzer — applying Generative AI to a practical enterprise IT operations use case. The solution analyzes incident data and provides insights into ticket patterns, complexity, root causes, and potential preventive actions. I built an interactive analytics experience using Python and Streamlit, integrating Claude to generate AI-powered incident insights. What made this project especially valuable was connecting GenAI concepts with a familiar enterprise delivery challenge — moving beyond dashboards and simply reporting what happened toward helping teams understand why it happened and what actions could help prevent recurrence. After many years of leading technology programs and transformation initiatives, I’m enjoying getting hands-on with GenAI and exploring how these capabilities can be applied to solve real enterprise problems and support better decision-making. A big thank you to The Gen Academy, and especially Aishwarya Srinivasan and Arvind Narayanamurthy, for the structured learning, practical guidance, and hands-on approach throughout this journey. 🙏 More projects and learnings to come! #GenerativeAI #GenAI #ArtificialIntelligence #AITransformation #TechnologyLeadership #DigitalTransformation #Python #Streamlit

As I wrap up this learning journey, I can clearly see the difference between before joining this cohort and now. - Through the connections I’ve built in this cohort, the quality of content I come across on LinkedIn has improved immensely. - New AI buzzwords no longer feel completely unfamiliar. I can understand the context without spending hours going back to the fundamentals. - Most importantly, I’m able to place new concepts into my mental map — where they fit, what they solve, and where they can be useful. A recent example is Jev and the discussion around the “Internet moment for AI.” Being able to follow the conversation, understand it, and relate it to what I’ve learned was a small but meaningful realization for me. And with every new wave of AI buzz, I still look to Aishwarya Srinivasan, Arvind Narayanamurthy and The Gen Academy to help cut through the noise and demystify it. A sincere thank you to Aish and Arvind for the knowledge, perspective, and confidence to keep learning and connecting the dots. 🙏 A special thanks to Aswath Harish Jayaprakash for bringing so many of us from the cohort together and helping build connections beyond this. And to Shanly Rajan, for being an excellent project mate throughout this journey. 🙌 #AI #AgenticAI #GenAI #AILearning

Happy to share that I’ve completed the Mastering Agentic AI course on Maven with The Gen Academy. 🚀 It’s been a great six-week journey of learning, building, and exploring topics across RAG, AI agents, MCP, evaluations, fine-tuning, multi-agent systems, and AI security. It was a great experience learning from Aishwarya Srinivasan and Arvind Narayanamurthy throughout the cohort and gaining practical exposure to Agentic AI. Looking forward to applying everything I learned and continuing to build in Agentic AI. #AgenticAI #GenerativeAI #AIEngineering #TheGenAcademy #Maven #BuildInPublic #LearnInPublic

Six weeks into Agentic AI, and I’m noticing a change I didn’t expect. 🚀 I recently completed the Mastering Agentic AI course on Maven with The Gen Academy. We covered RAG, AI agents, MCP, evaluations, fine-tuning, multi-agent systems, and AI security—and spent time building, not just listening. I’m proud to have completed the course. But the biggest change is how I think about AI now. 🔹 I have a better signal. The people I connected with in the cohort have changed the quality of AI content and conversations I see on LinkedIn. 🔹 The buzzwords make sense. When a new term shows up, I can usually follow the conversation without starting from the fundamentals again. 🔹 I have a mental map. I can place a concept in the larger architecture: what it solves, where it fits, and when it’s useful. That became real when my feed filled with discussions about MCP, agent evaluations, and multi-agent orchestration. I could follow the debate—and connect it to what I’d built during the course. It sounds small, but that was a meaningful moment for me. There will always be another wave of AI hype. I’m grateful to Aishwarya Srinivasan, Arvind Narayanamurthy, and The Gen Academy for helping me find the substance underneath it—and for giving me the confidence to keep connecting the dots. 🙏 Now the real work begins: keep applying what I’ve learned and keep building. For the first time, I can see how the pieces fit together instead of trying to keep up with every new AI trend. Upward and onward. 🚀 #AgenticAI #GenerativeAI #AIEngineering #TheGenAcademy #Maven #BuildInPublic #LearnInPublic

Six weeks of Mastering Agentic AI with The Gen Academy on Maven have strengthened my ability to connect AI concepts, architectural decisions, and working implementations. I entered with more than a decade in quality engineering and test architecture, experience in AWS and CI/CD, and early practical familiarity with assembling AI workflows. My objective was to bring that engineering discipline into AI system design: understanding how models, context, tools, orchestration, and controls work together—and being able to defend the choices. The program connected six areas of practice: • Model foundations — capabilities, context limits, structured outputs, function calling, and quality, latency, and cost trade-offs. • RAG and context engineering — chunking, embeddings, vector databases, retrieval strategies, and grounding. • Agentic architecture — workflow and multi-agent patterns, state management, tool integration, MCP, and LangGraph orchestration. • Model adaptation — choosing among prompting, retrieval, and fine-tuning, including LoRA/QLoRA and local inference. • Evaluation and observability — golden datasets, human and subject-matter review, model-based judges, traces, and regression monitoring. • Security and governance — prompt injection, sensitive-data handling, permissions, guardrails, red teaming, and runtime controls. I applied those concepts in project work: • Built a multi-tenant RAG system with grounding, provenance, tenant isolation, and a separate evaluation corpus. • Developed a bounded financial-document reviewer using LangGraph to coordinate extraction, verification, and critique through typed handoffs, evidence checks, controlled tools, and a deterministic final gate. • Built ClaimLens, applying a fine-tuned vision model to damage assessment. • Extended the financial-document reviewer with system-level evaluation and security testing mapped to the OWASP Top 10 for LLM Applications. Harness engineering connects these decisions. The runtime harness manages context, tools, state, permissions, and execution. The evaluation harness examines both outcomes and the behavior that produced them, including handoffs, tool use, failure handling, and operational constraints. The most valuable outcome is greater clarity about architectural trade-offs: where reasoning adds value, when retrieval or adaptation is warranted, when multiple agents justify their coordination overhead, and where deterministic controls and human review belong. Thank you Aishwarya Srinivasan and Arvind Narayanamurthy for the technical depth, practical insight, and thoughtful structure you brought to The Gen Academy's six-week program on Maven—and for encouraging us to build, test, and challenge our architectural decisions. #AgenticAI #AIEngineering #AIArchitecture #LangGraph #GenerativeAI

Hi Bay Area folks! 👋 I’m attending this event that Aishwarya Srinivasan is hosting with her amazing co-host Arvind and I’m genuinely so excited for it! I’ve been watching these two pour so much time, energy, and heart into their sessions and courses at GEN Academy, and it’s inspiring to see how passionate they are about educating people and building a strong community. I can’t wait to meet them in person and soak up some of that contagious energy and love for the community that I’ve been seeing through LinkedIn. ❤️ This is going to be such a fun evening! If you’re joining, ping me or come say hi. Would love to meet you there! ✨

AI is exciting. But for me, the real question is: can we use it to make someone’s day a bit easier? I had the chance to attend the /build-with-AI Buildathon at Carnegie Mellon University, hosted by The Gen Academy in partnership with Replit and CMU Heinz College. Our project focused on supporting people living with dementia and Alzheimer’s. We built Recollect, a simple daily companion designed to help patients understand what is happening right now, what comes next, and who they can turn to for help. Caregivers can create a daily plan, while patients see a simplified experience with reminders, familiar-person recognition, and clear next steps. The challenge wasn’t building more features or making the UI look impressive. It was figuring out how to make it simple and accessible enough for someone with dementia to actually use. That was probably my biggest takeaway from the buildathon. With AI, it is really easy to keep adding things just because we can. But for Recollect, we had to keep asking what was actually useful and what would just make things more confusing. Sometimes building a better product isn’t about adding more but knowing what to leave out. And a big thank you to Myra, Raasi, and Archana for building this with me, and to Aishwarya Srinivasan, Brandon Middleton, Arvind Narayanamurthy, Tanvir Ahmed Shaikh, and Amish Suchak for the feedback and support throughout the buildathon!

I’m happy to share that I’ve obtained a new certification: Mastering Agentic AI: Certification by The Gen Academy from Maven! Massive thanks, Aishwarya Srinivasan and Arvind Narayanamurthy, for sharing your expertise and pushing us from theory to real-world implementation.

As a Quality Engineer and Business Analyst in Test, I spend a lot of time translating requirements into testable scenarios. So, when our Week 5 assignment focused on fine-tuning, I chose a question from my own work: could a small open-weight model turn user stories and acceptance criteria into useful test cases? I fine-tuned Qwen3-1.7B with LoRA to generate structured positive, negative, and boundary tests. Each test included preconditions, steps, expected results, and traceability to the acceptance criterion it covered. I used LLaMA Factory and Google Colab to train and evaluate the model. My first fine-tuning attempt didn’t improve the model. A clearer prompt helped more, which made me question whether fine-tuning was adding value at all. Then I did what testers do: I questioned the test data. The dataset contained predictable patterns, allowing the model to take shortcuts instead of understanding the intent behind each criterion. I rebuilt it with 440 varied examples, shuffled the acceptance criteria, and separated scenario families for training and final testing. On 30 held-out examples from three unseen scenario families, the fine-tuned model outperformed the base model: Test-type accuracy: 62.8% → 90.8% Overall score: 83.4% → 91.3% This reinforced something familiar from testing: the quality of an outcome depends heavily on the quality of the examples, rules, and evaluation criteria. I see tools like this supporting testers and business analysts by drafting test cases, improving requirements traceability, and helping uncover scenarios that need further discussion. Human review still matters for business context, risk, and edge cases. For QAs wondering where they fit in AI work: evaluation is one area where our skills matter a great deal. Big thank you to Aishwarya Srinivasan and Arvind Narayanamurthy for making these complex concepts easier to understand and encouraging us to apply them to problems from our own work! Repo: lnkd.in/gpFhPEW4

Great learning session on GenAI Governance, Security, and Observability! I had the opportunity to attend an insightful presentation that explored some of the real-world challenges organizations need to think about as they move GenAI applications from experimentation into production. One topic that particularly stood out to me was how LLM/GenAI observability and evaluation can be aligned with the OWASP Top 10 risks for GenAI, including: 🔹 Prompt Injection 🔹 Sensitive Information Disclosure 🔹 Data/Model Poisoning 🔹 Improper Output Handling 🔹 Excessive Agency 🔹 System Prompt Leakage 🔹 Vector/Embedding Weakness 🔹 Misinformation & Hallucinations As we build more RAG systems and AI Agents, it is becoming increasingly clear that getting a good response from an LLM is only part of the solution. We also need to think about evaluation, security, governance, monitoring, and trust throughout the AI application lifecycle. Thank you Aishwarya Srinivasan and Arvind Narayanamurthy for putting together this thoughtful session. Also appreciate Nick Nolan, for the insightful discussion and knowledge sharing. Great session and lots of valuable takeaways! Aadhil A. Majeed John Ninan Aswath Harish Jayaprakash Gautham Guganesh Avantika Dharmagadda, PMP #GenerativeAI #GenAI #AIAgents #RAG #LLM #AIGovernance #AISecurity #LLMOps #OWASP #ResponsibleAI

This week, I built AI workflows in two completely different ways — with code and with no-code. 🔧 Same goal: build AI that actually does useful work. Two very different approaches. CODE → FinRAG I built a RAG system in Python using LangChain that answers questions over real SEC filings. → Custom chunking → Reranking → Retrieval strategy → Full evaluation harness It gave me deep control over every part of the pipeline. NO-CODE → n8n I built an AI workflow visually by connecting a trigger, an AI Agent, and a Gemini model. Input: a paragraph. Output: a clean, structured summary. From idea to working automation in minutes. Building both back-to-back made one thing very clear: Agent concepts make much more sense when you see them working. In n8n, memory and tools aren't just abstract concepts — they're components you connect when an agent needs to remember something or take an action. And the biggest lesson? Code gives you control. No-code gives you speed. FinRAG needed control because retrieval, chunking, reranking, and evaluation required deeper customization. The n8n workflow needed speed — and no-code was perfect for that. They aren't competitors. They're different tools for different jobs. The real skill is knowing which one to reach for — and when. Grateful to Aishwarya Srinivasan and Arvind Narayanamurthy for continuing to push us to build across the entire AI stack, not just the parts we're already comfortable with. #AgenticAI #n8n #RAG #LangChain #AI #NoCode #BuildInPublic #TheGenAcademy

InterviewDNA — a personalized interview coach that reads your resume and a target job description, finds the real gaps, and runs an adaptive mock interview that changes in real time based on how you actually answer. Not a chatbot. Not a static question bank. An actual agentic system. **What that meant in practice:** → A LangGraph state machine — not the LLM — decides what happens after every answer: probe deeper, raise the difficulty, coach a real knowledge gap, or move on. The model reasons and generates; the code decides and controls. → Hybrid search for grounding — vector search (Pinecone) + keyword search (BM25) fused together, then reranked with a cross-encoder. Because semantic search alone misses things a plain keyword match would catch, and vice versa. → A genuine tool-use agent — when it detects a knowledge gap, it decides whether to search its own indexed material or the live web, then grounds its coaching in whatever it actually finds. No hallucinated explanations. → An evals harness, not just unit tests — because "does my code run" and "is the model's output actually good" are completely different questions. I built LLM-as-judge checks specifically to catch things keyword matching can't: hallucination, unfaithful grounding, subtle scoring bugs. I chased a production bug for hours that turned out to be blocking async code freezing an entire FastAPI server. I hit a hard memory ceiling on free-tier hosting and had to actually understand what PyTorch's CUDA-bundled wheel was doing to my container's RAM. I found a race condition between writing to a vector database and immediately querying it. I debugged a model provider silently deprecating the exact model I'd built against. Every one of those was a systems engineering problem wearing an AI costume. That's the actual skill gap between "using AI" and "engineering AI systems" — and it's a lot closer to backend engineering than the hype suggests. Fully self-hosted, fully free to run end to end: local LLM, hybrid retrieval, structured evals, real deployment debugging included. Happy to talk through any piece of this — the agentic loop, the retrieval design, or the very unglamorous production bugs that taught me the most. github: lnkd.in/gvJuZq8n Thanks to Aishwarya Srinivasan and Arvind Narayanamurthy pushing hands-on building over passive learning. #AIEngineering #LangGraph #RAG #BuildInPublic

Week 1 of Mastering Agentic AI started this weekend, a six-week cohort on Maven taught by Aishwarya Srinivasan and Arvind Narayanamurthy through The Gen Academy. The first session had over 350 people on the call. The chat filled with locations as people joined, North America, India, Europe, Australia, Singapore, South Africa, Dubai. One person was grabbing coffee at 1am local time to attend live! The sheer range of people deciding agentic AI is worth their Saturday says something before any curriculum does. I've followed Aishwarya's work for a while, and her approach to teaching AI, grounded in building rather than theory, is a big part of why I signed up. The program runs live sessions, guest lectures across time zones, and group builds. I'll be sharing what I learn and what I build as the six weeks go on, as the reading list alone has already changed what my browser tabs look like. #AgenticAI #AI #BuildingInPublic

Starting a New Learning Journey: Mastering Agentic AI: Certification by The Gen Academy Excited to start the Mastering AI Agents 6 Weeks cohort with Aishwarya Srinivasan , Arvind Narayanamurthy and The Gen Academy !!! As I continue expanding my journey from Data Engineering into AI/ML and Agentic AI, I’m looking forward to moving beyond understanding individual AI concepts toward learning how to design and build production-grade AI applications. We kicked things off with a strong foundation, covering: 🔹 Evolution of LLMs 🔹 Determinism vs. Non-Determinism 🔹 LLM Architecture & Model Landscape 🔹 Prompt Foundations & Techniques 🔹 AI Agents, Harnesses and Loops 🔹 The growing Agentic AI technology stack One thing is already clear: this is going to be a marathon, not a sprint. 🏃♂️ There is a huge difference between experimenting with an LLM and building an AI application that is reliable, observable, secure, and production-ready. I’m especially excited to learn how all the pieces of the Agentic AI stack come together to bridge that gap. Looking forward to learning, experimenting, and most importantly, building production-grade applications alongside this cohort over the coming weeks. The journey continues! 🚀 John Ninan Gnana Sudheer Gavarraju Aswath Harish Jayaprakash Gautham Guganesh #AgenticAI #AIAgents #GenerativeAI #LLM #ArtificialIntelligence #PromptEngineering #AIEngineering #MachineLearning #DataEngineering #ContinuousLearning

Yesterday, I embarked on the marathon learning journey of Mastering Agentic AI with The Gen Academy, led by Aishwarya Srinivasan and Arvind Narayanamurthy. Coming from a predominantly business and transformation background rather than a deep technical one, I was initially curious about how accessible this learning experience would be. What stood out immediately was Aishwarya's exceptional ability to simplify complex AI concepts into practical, easy-to-understand insights. The introduction to vibe coding was particularly inspiring, making AI development feel far more approachable and motivating than I had imagined. I'm excited about the journey ahead and look forward to expanding my understanding of Agentic AI, building hands-on skills, and exploring the possibilities this technology creates. Thank you, Aishwarya and Aravind, for creating such an engaging and inclusive learning experience. Looking forward to learning much more over the coming weeks! #AgenticAI #GenerativeAI #AITransformation #ContinuousLearning #GenAcademy #Aishwarya #Aravind

This post was drafted by an agent I built this week. 🤖 Week 3 of the Mastering Agentic AI bootcamp, and I shipped my first end-to-end agentic system: a multi-agent pipeline with a human in the loop. The problem was personal. I want to keep building in public, but writing a post every week eats 2–3 hours of re-reading slides and staring at a blank page. So I made that the project. Meet the Build-in-Public Content Agent. It ingests a week's course materials, my own notes, and my project repo, then pairs what I learned with what I actually built. A planner picks the themes, a generator drafts for LinkedIn and my GitHub Pages blog, and a critic scores every draft for grounding and tone — looping back until it clears the bar. Nothing publishes without me. A LangGraph interrupt() halts the graph and hands me the draft. Reads are autonomous; writes wait for a human. The lesson that stuck: the hard part of agents isn't the prompts. It's control flow, state, and drawing the line between what the agent does alone and what needs a human. (Yes — it wrote this too.) Code + a 5-min demo in the comments 👇 What weekly workflow would you turn into an agent first? Huge thanks to Aishwarya Srinivasan and Arvind Narayanamurthy for creating a course that just gets better every week! #AgenticAI #LangGraph #BuildInPublic #LLM

🚀 Week 3 Project Done — Bulls & Cows meets LangGraph! For this week's project, I built an Agent vs Human Bulls and Cows game powered by LangGraph and Streamlit — and I'm genuinely proud of how it came together. 🎯 What it does: An AI Opponent Agent uses LangGraph to intelligently guess your secret number, while a Coach Agent watches every move, tracks the full game history — guesses, feedback, candidate counts — and gives you real-time reasoning notes to help you play better. Both agents can be powered by a local Ollama model or hosted LLMs like Gemini or Nebius. Under the hood: deterministic tools handle the bulls/cows scoring and candidate filtering so the game stays correct, while the LLM makes it conversational and engaging. LangSmith tracing ties it all together for visibility into every agent turn. 🔧 Stack: Python · LangGraph · Streamlit · Ollama · LangSmith · Nebius/Gemini 👉 lnkd.in/eCea8efg Huge thanks to Aishwarya Srinivasan and Arvind Narayanamurthy — your teaching pushed me to think in graphs and agents, not just scripts. The way you broke down LangGraph concepts week by week made this project feel achievable. And to my amazing cohort — watching your projects, your questions in sessions, and your encouragement kept me going when the agent wasn't cooperating 😅. You all set the bar high and inspired me to clear it. On to Week 4! 💪 #LangGraph #AI #AgentAI #Python #Streamlit #LLM #BuildingInPublic #Week3

Built an AI-Powered Jira Sprint Reporting Agent this week. 🚀 --------------------------------------------------------- The goal was simple: Instead of manually reviewing Jira boards every week, let an agent fetch sprint data, analyze project health, identify blockers, and generate an executive-ready status report. How it works ** Connects directly to Jira Cloud using the latest Atlassian REST APIs ** Fetches: - Active sprint - Sprint issues - Issue details - JQL search results Historical sprint information 🔹 Normalizes and enriches the data 🔹 Uses an LLM as the reasoning engine to: Summarize sprint progress 1. Identify blockers and delivery risks 2. Highlight sprint trends 3. Recommend actionable next steps 4. Predict areas that need attention in the next sprint The interesting part 👇 The agent doesn't stop after generating a report. It stores weekly snapshots (metrics, blockers, risks, recommendations) and validates itself when the next sprint starts. That means it can answer questions like: ✅ Which issues have been blocked for more than one sprint? ✅ Were the predicted risks actually realized? ✅ Did blocker resolution improve sprint velocity? ✅ How accurate were last week's recommendations? One takeaway from building this: Not everything needs AI. Fetching data, calculating velocity, identifying blockers, and comparing sprints can all be implemented with deterministic business rules. The LLM becomes valuable when it's time to transform structured project data into executive summaries, explain risks, recommend actions, and answer natural language questions. That's where AI complements analytics rather than replacing it. Curious to hear how others are approaching AI-powered project intelligence beyond traditional dashboards. Aishwarya Arvind Repo: lnkd.in/gaFBMK7B #LangGraph #LangChain #Jira #Atlassian #Python #LLM #ProjectManagement #Automation

Is your AI actually remembering anything, or is it just repeating itself? In B2B work, 'memory' isn't just about recalling facts—it’s about preventing context from poisoning the next decision. #GenAcademy had Rudraj Mehta (Product Lead at #Mem0) walk through how memory algorithms actually work — the session demo was a restaurant chatbot that remembers you like pasta and avoids mushrooms. Simple example, makes the architecture click. But the three phases he described — extract, consolidate, retrieve — are exactly what #B2BAI teams keep rebuilding from scratch when their agents lose context between sales calls. The things #RAG and Claude alone don't cover in a cross-session B2B context: No persistence between sessions. Your CRM agent meets every prospect cold, every time. No deduplication or contradiction management. A prospect's budget moved from $80K to $120K across three sessions. Which number does your agent use in the fourth? No entity graph. Knowing Sarah is the engineering veto, she's at TechVision, and TechVision is expanding to APAC in Q3 — those are relationships that need to be traversable, not flat stored text. Then the Q&A happened. A builder asked what nobody writes about in courses: what if the same user ID carries memories from two contexts that must never cross-contaminate — a seller and a buyer on the same platform, or a patient and a caregiver in healthcare? One memory set poisoning the other. Rudraj's answer was honest: "That's an interesting edge case. There's no dynamic way to deal with this right now." And then he walked through how to think about scoping it manually — namespace isolation at the Run ID level, context-specific extraction instructions, promoting shared signals deliberately. That answer is more useful than most content you'll find in any AI course. It's what happens when real builders push the edges of a real system live. Credit to Rudraj Mehta for the transparency, and to Aishwarya Srinivasan, Arvind Narayanamurthy for creating a format where that kind of question gets asked and answered in public. Full B2B walkthrough with an interactive example here: lnkd.in/gdHk4WNm

Week 3 : I built a multi-agent tax-research system that answers questions about current 401(k) and retirement rules . 🔹 A different model for every agent A Claude Sonnet orchestrator decomposes the question. Fast, cheap Haiku researchers run search in parallel. GPT-4o-mini extracts structured claims. The result: frontier reasoning where it matters, cheap horsepower where it doesn't. 🔹 Decorrelated errors as a feature. The verifier runs on a different model family than the extractor that produced the claims , so it catches mistakes a same-model check would happily rubber-stamp. The system actually enforces this at startup. 🔹 Ground truth beats vibes. A deterministic knowledge base of curated IRS figures is the verifier's source of truth. Every claim carries a source URL, snippet, and date. The synthesizer can only cite supported, authoritative claims ,and it explains the rules without crossing into personalized financial advice. 🔹 Built on LangGraph with parallel fan-out and a bounded orchestration loop that sends the graph back to re-plan when it spots a gap. The big lesson: trustworthy AI isn't about one giant model knowing everything. It's about architecture , routing the right model to the right job, and engineering in disagreement so errors surface instead of slipping through. Thanks to Aishwarya Srinivasan and Arvind Narayanamurthy :) #AI #MultiAgent #LangGraph #LLM #MachineLearning #AIEngineering lnkd.in/giTt-TSs

📆 Week 3 - The Gen Academy ⚽ World Cup fever has taken over Toronto right now — the city is absolutely electric. 🏆 So this week I built something that mixes the two things on my mind: the beautiful game, and agentic AI. Meet #VirtualSoccerArena — #Week3 of the Agentic AI Bootcamp. It's an MCP server that lets you book a soccer pitch just by chatting with Claude. Ask "what's open June 28, what's the weather, and what's the rain policy?" — and Claude pulls the forecast, checks live availability, answers the policy question from a 100-page rulebook via RAG, and books your slot. Four tools, one ReAct loop, one natural conversation. A web app then reads the same booking and unlocks a goalkeeper mini-game. 🎮 Playful on the outside. But under the hood, it's a masterclass in what production agentic AI demands: 🧠 One brain, two front doors — Claude and the web app share a single core and a single source of truth. Zero drift. 🔒 Human-in-the-loop where it counts — Claude reasons and drafts on its own, but every irreversible action (booking, email) waits for a human yes. That's the line between a helpful agent and a liability. 🗄️ The right tool for the data — SQLite for transactional bookings, LlamaIndex + RAG for the policy PDF. Match the store to the shape, not the hype. The takeaway that'll stay with me: agentic AI is really a distributed-systems problem wearing a friendly chat interface. MCP gives you a clean integration contract, shared state keeps your front doors honest, RAG grounds the agent in real knowledge, and human-in-the-loop gates keep it safe — but it all only holds together because the logic lives in one core and one source of truth. The model is the exciting part; the architecture is what earns the trust. And it scales beyond soccer — the same Quote-to-Booking blueprint fits banking, telecom, healthcare, oil & gas, and e-commerce. Kudos to Aishwarya Srinivasan and Arvind Narayanamurthy for curating projects that don't just teach the theory, but force you to stress-test it. Building these systems from the ground up is the only way to truly internalize the concepts. 👏 Game on. 🔗 Code: lnkd.in/gKEeEKpN Slides below 👇 #AgenticAI #MCP #AIArchitecture #WorldCup2026 #SolutionArchitecture #RAG

MCP is getting more production-ready. The biggest shift I’m watching is not “more tools.” It is a better infrastructure for remote agent systems. The latest MCP spec keeps the familiar initialization flow, but adds important production pieces: - Streamable HTTP as the main remote transport - optional session management with MCP-Session-Id - polling-friendly SSE behavior - resumability with Last-Event-ID - stronger authorization discovery with OIDC support - experimental tasks for durable requests - improved elicitation for human-in-the-loop workflows The takeaway for AI engineers: MCP is moving from “local connector demo” toward a real application layer for agentic systems. That matters because production agents need more than tool calls. They need: - secure authorization - resumable execution - clear user consent - human approval flows - durable task tracking - infrastructure that survives real network conditions I am learning about this in detail, thanks to the Cohort I am enrolled in with great mentors like Aishwarya Srinivasan and Arvind Narayanamurthy I’m especially interested in how teams will manage state across remote MCP servers. Do you keep state in the server session? Push it into durable task storage? Or design the agent loop to be mostly stateless and resumable? This feels like one of the big architecture questions for agentic AI. #MCP #AgenticAI #AIEngineering #SystemArchitecture #CloudComputing #SoftwareEngineering #BuildInPublic

💵 Currency conversion tells you what something costs. It doesn't tell you whether it's actually affordable. That's the problem I explored in my latest project: 𝐆𝐥𝐨𝐛𝐚𝐥 𝐂𝐨𝐬𝐭 𝐀𝐝𝐯𝐢𝐬𝐨𝐫. The platform combines purchasing power parity, cost-of-living data, and AI-powered insights to help travelers, international students, digital nomads, and relocating professionals understand what prices actually mean across different economies. For example, a $6 coffee in San Francisco may seem expensive when converted to another currency, but is it actually expensive relative to local purchasing power? Building this project reinforced something I've been learning about AI: The value isn't in generating answers. It's in helping people make better decisions. GitHub: lnkd.in/gUASM2ny Thank you to Aishwarya Srinivasan and Arvind Narayanamurthy as I continue learning and building in AI, their content, guidance, and practical approach have been incredibly valuable in helping me. What other everyday decisions could AI help people make better? #AI #GenerativeAI #ProductManagement #BuildInPublic #ArtificialIntelligence

The term "agentic AI" is getting stretched to cover three very different things. Worth untangling, especially if you're scoping your own build. This is something Arvind Narayanamurthy and Aishwarya Srinivasan touched on earlier in the cohort, and it's been resonating with me ever since. Take a simple use case: "help me stay under $500 for groceries and dining out this month." 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗼𝗻 runs a fixed sequence, tracks spending, and sends an alert when it crosses $500. No reasoning, no adaptation. Reliable as a Burglar alarm, but it only tells you after you've already blown the budget. 𝗥𝗔𝗚 adds reasoning over retrieved context: it pulls your transaction history and spending patterns, and generates a recommendation grounded in both: "at this pace, you'll hit $500 by the 22nd." One pass → retrieve, reason, respond. Smarter, but it stops there. It told you the problem; it didn't do anything about it. 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 closes the loop. The system spots the overspend risk, decides on its own to dig into which subcategory is driving it, proposes pausing two unused subscriptions, checks if that's enough, and if it isn't → then proposes shifting funds from a category with slack, checks again, and only stops once it's actually found a path under $500 (or reports back honestly if it can't). That repeated act → check → act-again cycle, happening without you re-prompting at each step, is what makes it agentic. Remove that loop and leave just one shot, and it's not agentic anymore, just RAG with extra confidence. A natural follow-up: what if the sub-tasks ran in parallel instead of one after another? That's 𝗺𝘂𝗹𝘁𝗶-𝗮𝗴𝗲𝗻𝘁: a spend-analysis agent, a subscription agent, and a reallocation agent each working their angle independently, with an orchestrator merging the results into one recommendation. Sounds more sophisticated. Often isn't worth it. Multi-agent earns its complexity when tasks are genuinely independent and expensive enough that parallelizing actually saves time. For one month's grocery budget, a single agent looping through the same checks sequentially gets you there just as well → faster, and with far less to debug. The coordination overhead of multiple agents (reconciling outputs, more API calls, more failure points) isn't free. The real test, regardless of architecture: if you removed the loop, would the system behave the same way? If yes, it was never agentic to begin with, just a single LLM call wearing the label. And the same test applies one level up: if you removed an agent, would the outcome change? That's the harder question I've been sitting with on my own work.

Week 2 of our Mastering Agentic AI Certification, May Cohort is officially a wrap! Huge congratulations to Santosh, our AI Quiz Wiz of the Week! Arvind and Aishwarya are incredibly proud of what they're building with The Gen Academy. The engagement, curiosity, and energy from this cohort has been something truly special to witness. One thing they care deeply about is live learning. Learning AI is not just about consuming information. It is about asking questions, challenging ideas, learning alongside others, and building intuition through discussion. The energy that comes from a live classroom is something recorded lessons simply cannot replicate. That is why every session includes live Q&A, open discussions, and weekly quizzes that make learning interactive and fun. This week, Santosh earned the title of AI Quiz Wiz of the Week. Well deserved! P.S. See that "Artificially Intelligent" hat at the top? Santosh is officially getting one. They take quiz victories very seriously around here. Want to be part of a cohort like this? The August Cohort for the Mastering Agentic AI Certification is now open. Live sessions, real discussions, weekly quizzes, and a community that actually shows up for each other. 👇 Link in the comments — come join us.

Attended a guest lecture by Arjun Patel, Senior Developer Advocate at Pinecone, on retrieval patterns for RAG pipelines. My head is full of ideas. Here are my biggest takeaways: 1. How you store your data = how you can query it This sounds obvious until you actually build a RAG system. The embedding model you choose, how you chunk your data, and which index structure you use all shape the kinds of retrieval you can support later. Make those decisions intentionally. 2. Retrieval is not one thing Arjun walked through a live bird search demo covering: - Dense semantic search for vocabulary mismatch - Full-text / keyword search for precision queries and rare terms - Visual/image search using shared text-image vector space - Hybrid + filter for compound queries - Coarse-to-fine retrieval by filtering first, then searching inside the narrowed context - Agentic search, where Claude decides which tool to call based on query type 3. Evals do not have to be scary You do not need 1,000 labeled examples to start. Start with 5 to 6 queries you care about. Define what a good result looks like. Check whether your retrieval is actually improving. Diagnose first. Scale later. The mental model I am taking away: Improving retrieval improves RAG. Everything else is downstream of this. I am already applying these ideas to CohortBrain, a GraphRAG + Vector RAG knowledge system I am building for my bootcamp. The timing of this lecture could not have been better. If you are building RAG systems and not thinking deeply about your retrieval layer, this is your sign. Thanks to Arjun Patel and Aishwarya Srinivasan , Arvind Narayanamurthy for making this happen. #RAG #VectorSearch #Pinecone #LangChain #AIEngineering #BuildInPublic #GenerativeAI

For Week 2 of my Agentic AI cohort, I built a Rare Disease Workup Assistant — an 𝐀𝐈-𝐩𝐨𝐰𝐞𝐫𝐞𝐝 𝐜𝐥𝐢𝐧𝐢𝐜𝐚𝐥 𝐝𝐞𝐜𝐢𝐬𝐢𝐨𝐧 𝐬𝐮𝐩𝐩𝐨𝐫𝐭 𝐭𝐨𝐨𝐥 that helps primary care physicians identify potential rare diseases from free-text patient descriptions in under 30 seconds. 𝐓𝐡𝐞 𝐩𝐫𝐨𝐛𝐥𝐞𝐦 𝐰𝐨𝐫𝐭𝐡 𝐬𝐨𝐥𝐯𝐢𝐧𝐠 300 million people worldwide live with a rare disease. The average diagnostic odyssey takes 4–6 years of misdiagnoses, unnecessary treatments, and unanswered questions for patients and families. Primary care physicians are the first door most of them walk through. Expected to recognize patterns across 7,000+ conditions, with 15-minute appointment slots and limited specialist access. That gap — between what physicians are asked to do and what's humanly possible — is where AI can make a difference. Not by replacing clinical judgment. By making sure the right question gets asked before the patient leaves the room. 𝐖𝐡𝐚𝐭 𝐈 𝐛𝐮𝐢𝐥𝐭 A multi-stage Hybrid RAG pipeline — agentic phenotype expansion, BM25 + vector retrieval via Qdrant, intelligent reranking, and grounded generation via Claude Sonnet. All orchestrated through n8n. 𝐓𝐡𝐞 𝐞𝐯𝐚𝐥 𝐫𝐞𝐬𝐮𝐥𝐭𝐬 ✅ 95.7% Top-1 retrieval accuracy ✅ 83% Faithfulness ✅ 95% Answer Relevance Strong numbers. But the most instructive result wasn't one of these. The result that made me stop and think 71% 𝐬𝐩𝐮𝐫𝐢𝐨𝐮𝐬 𝐜𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐜𝐞 𝐫𝐚𝐭𝐞 on negative controls — the system flagged a potential rare disease 5 out of 7 times when presented with common condition descriptions. First instinct: that's a problem. Second instinct — after sitting with the confusion matrix: it isn't. A false negative in rare disease workup means a missed diagnosis. Another year on the diagnostic odyssey. A patient who walked in with answers walking out without them. A false positive means a physician follows an extra diagnostic thread. Inconvenient. Not harmful. A system calibrated toward sensitivity "one that would rather flag too much than miss something real" is not a flawed system. It's a responsibly designed one. The next iteration needs a confidence calibration layer — a mechanism that can distinguish "this genuinely looks rare" from "I'm uncertain, so I'm flagging it." That's the next build. 𝐖𝐡𝐚𝐭 𝐭𝐡𝐢𝐬 𝐩𝐫𝐨𝐣𝐞𝐜𝐭 𝐭𝐚𝐮𝐠𝐡𝐭 𝐦𝐞 Getting a RAG system to retrieve accurately is a technical challenge. Getting one that a physician will actually trust — transparent sourcing, calibrated confidence, explainable outputs — is a design challenge. The eval numbers open the door. The thinking behind them is what builds the trust. Grateful to Aishwarya Srinivasan and Arvind Narayanamurthy at The Gen Academy for building a program that measures you not by what you learned, but by what you built and why. As Aishwarya Srinivasan says — "It's a marathon!" Week 3 starts now. 🏃‍♀️ #GenAI #RAG #HealthcareAI #RareDiseases #ClinicalAI #AIEvaluation #MasteringAI #TheGenAcademy

Another week, another guest lecture that changed how I think about building AI systems. This time, Logan Markewich , Head of Open Source at LlamaIndex, walked us through document parsing, retrieval design, and a live financial due diligence agent. Here is what stuck with me: 1. PDFs are a mess, and that is the whole problem PDFs were not built to be machine-readable. Data is scattered. Reading order is broken. Tables are chaos. LlamaIndex built both LlamaParse and their document parsing stack to help solve this: - fast local parsing - managed parsing APIs - chart understanding - semantic formatting - granular bounding boxes The better the parse, the better the downstream AI system. 2. You probably do not need embeddings yet This one hit different. Logan built a financial RAG agent with zero vector embeddings. Just keyword search over parsed documents. His take: For 6–50 docs, keyword search plus dynamic context windows can beat embeddings. Embeddings are for narrowing search space at scale. They are not always the best tool for precision. 3. The citation problem is harder than it looks The agent cited sources with bounding box highlights on the actual PDF page. Simple idea. Hard execution. Models subtly reword things. "$394,000" can become "394 billion." Now your exact match breaks. Logan’s solution was a waterfall of fallbacks: exact match → whitespace stripped → currency stripped → longest subtoken match And if all else fails, still surface the page number. Better than pretending the citation is perfect. 4. ParseBench: build your own benchmark LlamaIndex built their own eval set because existing academic benchmarks did not reflect what customers actually struggled with. They evaluated across: - table accuracy - content faithfulness - visual grounding - chart understanding - semantic formatting Sound familiar? It is the same mindset as retrieval evals: Start with the problems you care about, not only the benchmarks that already exist. I am using this thinking to sharpen the evaluation layer in CohortBrain, my GraphRAG + Vector RAG system for the bootcamp. The timing could not be better. Thanks to Logan Markewich and Aishwarya Srinivasan, Arvind Narayanamurthy for making these sessions happen. #LlamaIndex #RAG #DocumentParsing #AIEngineering #BuildInPublic #GenerativeAI #LLM

🚀 The Future of AI Isn't Bigger Models. It's AI Daemons. Over the past few weeks, I've been diving deep into research on AI agents, memory architectures, autonomous systems, and the emerging concept of AI Daemons. One thing has become increasingly clear: The future of AI is not just about building larger models. It's about creating persistent, memory-rich, autonomous systems that can continuously observe, reason, learn, and act. In traditional computing, a daemon is a background process that runs continuously without direct user interaction. In modern AI, this idea is evolving into intelligent agents capable of maintaining context, managing memory, planning actions, coordinating with other agents, and executing complex workflows autonomously. Recent research highlights several key developments: 🔹 MemGPT introduces the concept of treating Large Language Models like operating systems with hierarchical memory management. 🔹 Agentic AI architectures are moving toward structured systems consisting of perception, reasoning, planning, tool usage, collaboration, and execution layers. 🔹 Persistent memory frameworks such as Mem0 and soul.py are exploring how AI agents can maintain identity, continuity, and long-term memory across sessions. 🔹 Multi-agent frameworks including LangChain, LlamaIndex, CrewAI, and Swarm are enabling specialized agents to collaborate on increasingly sophisticated tasks. 🔹 Memory OS research proposes that memory itself should function as an independent operating layer for AI systems, separate from the reasoning model. The emerging architecture increasingly resembles: Perception → Memory → Planning → Action → Audit Instead of simply responding to prompts, future AI systems may continuously monitor environments, accumulate knowledge, adapt from experience, and execute decisions through coordinated networks of autonomous agents. What excites me most is that we're witnessing a shift from AI as a tool to AI as infrastructure. The academic and engineering communities appear to be converging on a common direction: The future of AI is not smarter standalone models. It is persistent, memory-driven, multi-agent daemon systems operating autonomously at scale. A special thanks to Yannic Kilcher for his exceptional research paper breakdowns and technical analyses. His content has been incredibly valuable in helping me understand complex AI concepts, agent architectures, memory systems, and emerging research directions. If you're interested in understanding AI research beyond surface-level tutorials, I highly recommend exploring his work. Also grateful to Aishwarya Srinivasan for her practical AI engineering content, which has helped bridge the gap between research papers and real-world implementation. As someone actively building AI projects and exploring agentic systems, this research has fundamentally changed how I think about the next generation of intelligent applications. What are your thoughts? Will AI Daemons become the foundation of future software systems? #ArtificialIntelligence #AIAgents #AgenticAI #GenerativeAI #MachineLearning #LLM #MemGPT #RAG #MemorySystems #MultiAgentSystems #AIEngineering #DataScience #FutureOfAI #AIResearch #CrewAI #LangChain #LlamaIndex #Innovation

When you deeply understand a system, you stop explaining it through terminology and start explaining it through intuition. That’s exactly what I saw Chris Alexiuk “The Wiz” from NVIDIA do while walking through a very complex RLVR training pipeline. Instead of starting with reward shaping, delayed credit assignment, multi-environment RL, reward hacking, Chris explained the system through things we humans already intuitively understand: 1. Models “want cookies” → reward optimization 2. SFT → like learning arithmetic before solving harder problems 3. SWE-focused RLVR → like project grading instead of pop quizzes because feedback arrives much later 4. Reward hacking → a robot trained to find yellow cheese starts chasing bananas instead What made these analogies powerful wasn’t that they simplified the system away. They preserved the actual dynamics: incentives, feedback loops, delayed rewards, curriculum learning, proxy optimization, and generalization pressure. That’s what strong technical intuition feels like. You stop describing the mechanics alone and start seeing the underlying adaptive system. Personally, I think this lesson extends beyond AI. Every system eventually learns the proxy you reward. Models do. Organizations do. People do. Every adaptive system eventually learns the proxy it is rewarded on.... Thanks to Aishwarya Srinivasan and Arvind Narayanamurthy for facilitating this awesome session! #AI #Nemotron #Nvidia

My take on a #GenAcademy guest lecture I attended today. First watch: lost me halfway through. Left thinking I just wasn't ready for the content. Second watch: started making sense. Third watch: 7 ideas I'd never seen in any course. The RCA on why I was slow: I was still thinking in traditional LLM mode. The lecture starts past that. If you're in the same spot, the blog I built links to my older posts on classical RL and LLM fundamentals — read those first, then come back. The 7 takeaways aren't academic. They're the kind of thing that only surfaces once you're actually building with LLMs: — Why RL needs SFT as a warm-up (and how little is enough) — How models learn shortcuts instead of goals — Why running 70+ environments at once isn't overhead — it's the fix — Why RLHF trained first produces a politely useless model Each concept has an interactive sandbox. lnkd.in/g22eJ-vv Thanks again to Chris Alexiuk for the lecture. Getting there took a few attempts, but it paid off. -------------------------------------- The irony isn't lost on me. Traditional LLM me rewatched the same lecture three times expecting different output — which is literally single-environment reward hacking. Same shortcut, same result, no progress. The fix, as I now know, was to change the environment: stop watching and start building. That's when the signal kicked in. Next session I'm showing up with better priors. SFT complete. Echoing what Aishwarya Srinivasan and Arvind Narayanamurthy mentioned about handling the complexity—it was tough, but making the mental model is what matters.

I was going through this slide in the Gen AI Building Blocks Bootcamp last week. 2026 Model Landscape. Open Source. Open Weights. Closed Source. And I realised — I had been using these terms loosely for years. Treating "open source" and "open weights" as if they meant the same thing. They don't. And if you're building anything that touches sensitive data, the difference is not academic. Let me explain it the way it finally clicked for me. ━━━━━━━━━━━━━━━━━━━━ Think of an AI model as a cake. ━━━━━━━━━━━━━━━━━━━━ The recipe = training code. Architecture, data, process — everything that created the model. The finished cake = the weights. Billions of numbers that ARE the model's intelligence. When it gives you a smart answer — those numbers are doing the work. Open Source → recipe AND cake. Bake your own from scratch. Full control. Examples: Llama 3.1, Mistral, Gemma 2. Open Weights → finished cake only. You can fine-tune it, run it, add your own flavours — but can't recreate it from scratch. Examples: Meta Llama 3.1 instruct, Cohere Command R+, Snowflake Arctic. Closed Source → you taste the cake through a window. Never hold it. Never see the recipe. Pay for every slice. Examples: GPT-4o, Claude, Gemini. ━━━━━━━━━━━━━━━━━━━━ What this means when you're building ━━━━━━━━━━━━━━━━━━━━ Closed source — fastest to start. Best reasoning. Just call the API. Perfect for prototyping and apps where privacy isn't a constraint. Open weights — right when data cannot leave your infrastructure. Medical records. Financial statements. PII. Run it on your server, data never touches a third-party API. Open source — for teams going deep. Reproduce training. Study architecture. Build derivatives. Essential for research. ━━━━━━━━━━━━━━━━━━━━ The mistake I see most often ━━━━━━━━━━━━━━━━━━━━ Teams assume "open" means private. It doesn't automatically. If you're calling someone else's hosted open-weights API — your data is still leaving your machine. The privacy benefit only kicks in when you self-host. Privacy isn't something you add later. It's a decision that shapes your architecture from day one. The cake is only yours if you bake it in your own kitchen. Next post — how to actually choose between these three when designing a real system. Three questions that make the decision obvious. Slide credit: Aishwarya Srinivasan and Arvind Narayanamurthy at The Gen AI Academy — Gen AI Building Blocks Bootcamp Week 1. #GenerativeAI #LLM #OpenSource #AIEngineering #PrivacyByDesign #HealthcareAI #MachineLearning #AIAgents #CloudSecurity #BuildingInPublic #AIBootcamp #ResponsibleAI

As part of the Mastering Agentic AI cohort led by Aishwarya Srinivasan and Arvind Narayanamurthy, there's been constant encouragement to move beyond learning concepts and actually build something useful. This is a small outcome of that push. I built a UK FTTP Opportunity Map using public Ofcom and ONS data. The dashboard helps explore: 1.Where full-fibre availability is strong 2.Where gaps still remain 3.Which local authorities have the most premises without FTTP 4.Oppurtunity score using coverage levels and premises without FTTP More than the dashboard itself, the project helped me think practically about how Agentic AI can support real analytical workflows — not just generating ideas, but shaping methodology, building the app, validating assumptions, documenting calculations, and getting it deployment-ready. Still early, still learning, but genuinely enjoying the process of building and experimenting. Grateful for the nudge to learn by doing. 🙏 🔗 Live app: lnkd.in/eSnXc75M 💻 GitHub: lnkd.in/eTwVVbQu

It's been close to a decade since I actually did some hands-on work in VS Code and GitHub. Courtesy of the Mastering Agentic AI course at The Gen Academy, I shipped my first vibe coded application an AI-Powered Product Description Generator (keeping things very simple!): a tool that helps eCommerce sellers generate fully optimized product listings. Here's what it does: 📝 Takes a product name, key features, an uploaded image, and a spec document as inputs 🛍️ Generates platform-specific listings for Amazon, Etsy, Shopify, and more 🔍 Produces SEO keywords — primary, secondary, and long-tail 🤖 Also generates some AEO and GEO content All powered by GPT-4o, built in React, and live on GitHub Pages. 👉 Try it here: lnkd.in/gP2bntEh What I learned building this: ✅ How to design and prompt multimodal AI API calls (text + image + document in a single request) ✅ How to structure prompts to enforce strict JSON output from LLMs ✅ How to go from an idea → a working prompt → a shipped app using vibe coding ✅ Version control and deployment via GitHub, GitHub Pages and Vercel A huge thank you to Aishwarya and Arvind for building a course that actually makes you ship things. :) I am in Week-1, so just getting started. More builds coming. 🔨 #AgenticAI #GenerativeAI #VibeCode #OpenAI #eCommerce #GenAcademy #MasteringAgenticAI #AIProducts #LearningInPublic

Week 1 of the Gen AI Building Blocks Bootcamp asked us to build something real. I built Finance-Management — an AI-powered personal finance analyser. The process taught me as much as the output did. WHAT IT DOES You upload your credit card statements — PDF, CSV, Excel, even a photo. The app reads them, categorises every transaction across 14 expense buckets, and gives you: → A financial health score (0–100) → Spend breakdown with 8 interactive Plotly charts → A personalised budget and savings plan → A chat interface where you ask questions about your own data → A downloadable PDF report + CSV export Across multiple currencies — USD, INR, GBP, EUR and 15 others. Never force-converts between them. ━━━━━━━━━━━━━━━━━━━━ 🔒 THE FEATURE I'M PROUDEST OF ━━━━━━━━━━━━━━━━━━━━ Before a single character reaches any AI model — the app runs a two-pass PHI masking pipeline. Pass 1 runs entirely offline on your machine: → Card numbers → [CARD-XXXX] → Account numbers → [ACCOUNT-XXXX] → Routing numbers → [ROUTING-XXXX] → IFSC codes → [IFSC-XXXX] → UPI IDs → [UPI-XXXX] → IBAN, SWIFT, CVV — all tokenised Pass 2 sends only the already-masked text to Groq — a single batched call — to catch anything regex missed. The user sees a green badge: "6 sensitive identifiers removed before sending to AI." Privacy isn't a checkbox. It's a design constraint that shapes everything else. STACK → Streamlit (4-tab UI: Upload / Dashboard / Plan / Chat) → UV for package management → Groq — llama-3.3-70b-versatile for analysis, plan & chat → Groq — llama-3.1-8b-instant for PHI validation & dashboard summary → Plotly + reportlab for charts and PDF export → Entirely on Groq free tier — no credit card, no paid models 🔗 lnkd.in/gaseXDXQ Repo includes full source, mock statements for testing, the vibe coding prompt I used, and setup instructions. Thank you to Aishwarya Srinivasan and Arvind Narayanamurthy at The Gen Academy The agentic mental map they taught — domain knowledge → goals → workflow → decisions → actions → autonomy — I used it before writing a single line of code. The PHI architecture, model routing, lazy-load pattern — all of it came from applying that framework to a real problem. That's what good teaching does. It gives you a way of thinking that travels with you. Week 2 is RAG. I'll keep sharing. #AIAgents #GenerativeAI #LLM #AgenticAI #Streamlit #Groq #LLMApps #HealthcareAI #CloudSecurity #FinTech #VibeCoding #AIBootcamp #Python #PrivacyByDesign #PersonalFinance

Closing out Week 1 of Mastering Agentic AI 🎯 and if I had to pick one key takeaway, it's this: "Attention Is All You Need" - the 2017 paper every LLM today is built on.📎 lnkd.in/gnsEPtMm I broke it down using the simplest possible example: "What is the capital of France?" The image below captures the full pipeline visually and here's what's actually happening step by step 👇 1️⃣ Splits the sentence into tokens — not words, chunks. "capital" becomes "cap" + "ital" 2️⃣ Converts each token into thousands of numbers that encode its meaning 3️⃣ Passes those numbers through stacked Transformer layers — at each layer, every token asks every other token: "how relevant are you to me?" 4️⃣ "capital" and "France" lock onto each other — not because they're next to each other, but because they're mathematically relevant. That's attention. 🔍 5️⃣ The answer builds one token at a time: "The"... "capital"... "of"... "France"... "is"... "Paris" 6️⃣ Tokens are assembled back into text and streamed to your screen The model doesn't read left to right like a human. It looks at every word simultaneously and scores relevance between every pair. That's why context window size matters. That's why prompt structure matters. You can't engineer what you don't understand. Week 2 goes deeper - RAG and context engineering are up next. Given everything this week surfaced, the timing couldn't be better. 🚀 Aishwarya Srinivasan Arvind Narayanamurthy #AgenticAI #MasteringAgenticAI #Transformers #LLM #RAG #ContextEngineering

I always wondered — when I type something to an AI, how does it actually understand me? Like, does it read words the way we do? Does it have some kind of internal dictionary? Turns out... it's something far more fascinating. And honestly, once I understood it, I couldn't stop thinking about it. Here's what actually happens under the hood 👇 𝟭. Your words get broken into tokens Before the AI reads anything, it slices your sentence into small chunks called tokens — roughly one word each. Every token gets a unique ID number. "How do planes fly?" becomes [4012, 278, 9141, 1823, 30]. Just numbers. 𝟮. Every word looks at every other word This is where it gets wild. The model doesn't read left to right like we do. Every single token scores every other token — asking "how relevant are you to understanding me right now?" This is called attention, and it's why the AI knows that "fly" matters more to "planes" than "the" does. 𝟯. This repeats across dozens of layers Each layer passes what it learned to the next — building from grammar → word relationships → meaning → intent. By the final layer, the model doesn't just know what your words mean. It knows what you're trying to ask. 𝟰. It writes the answer one word at a time The model predicts the single most likely next word. Then the next. Then the next — until a full, coherent response appears. No dictionary. No lookup table. Just patterns, numbers, and billions of learned weights. The GIF walks through all 4 steps visually — I made it because I couldn't find a simple enough explanation. If you've ever talked to ChatGPT, Claude, or any AI assistant — this is exactly what happened in the background, in milliseconds. Mind = 🤯 Aishwarya Srinivasan Arvind Narayanamurthy #ArtificialIntelligence #MachineLearning #AI #HowAIWorks #Tech #LLM

Just attended an incredible session on post-training and agentic AI by Chris Alexiuk from NVIDIA, hosted by Aishwarya Srinivasan and Arvind Narayanamurthy. My biggest takeaway? Post-training is where the magic happens. Pretraining gives a model knowledge. But post-training teaches it how to behave. That is the shift that took us from- Autocomplete machines → chat assistants → agentic models And now, systems like Nemotron-3 are pushing that even further. One idea from the session really stood out: A real agent is not just a chatbot. A real agent runs in a loop: goal + context → action → tool call → observe → update plan → repeat That loop matters just as much as the model itself. Because intelligence is not only about what the model knows. It is about what the system can reliably do. Another powerful idea: Reward can be a program. Instead of relying only on trained reward models, programmatic verifiers can check outputs using tools like: • pytest • SymPy • JSONSchema That makes reward signals cheaper, more auditable, and less likely to drift. This is huge for agentic systems. The infrastructure side was just as impressive. Scaling post-training is not only a research problem. It is an engineering problem. You need: • async rollouts • sandboxed environments • scalable evaluation • reliable verifiers • strong execution pipelines That is what makes agentic AI production-ready. The key message I took away: There is no wall in post-training. Better data, better environments, better algorithms, and better infrastructure can keep pushing model performance forward. The future of AI will not be built only by people who train bigger models. It will be built by people who understand how to shape model behavior after pretraining. Post-training is becoming the real battlefield. #LLM #ReinforcementLearning #NVIDIA #AgenticAI #PostTraining #MachineLearning #AI #DeepLearning

🏆 Can data predict the 2026 FIFA World Cup winner? I analyzed team-level data from six previous World Cups (2002–2022), alongside a separate 48-team prediction dataset for 2026. The data was sourced from Kaggle, and the complete analysis, predictive modeling, validation, and sports analytics dashboard were developed using Codex. 🔮 The model’s 2026 predictions: 🥇 Champion: Argentina 🥈 Runner-up: Germany 🔥 Semi-finalists: Argentina, Germany, Brazil and Spain 🎯 Other projected quarter-finalists: France, England, Uruguay 👀 Dark horses: Portugal, Netherlands The model considered: ⚽ Recent wins, losses and goal difference 📊 FIFA rankings and points 💰 Squad market value 🏟️ Host advantage 🏆 World Cup experience and knockout pedigree 🎂 Average squad age I also created an interactive-style Sports Analytics Dashboard covering team performance forecasts, advancement probabilities, tournament-stage classifications, and model validation. Of course, football is gloriously unpredictable, but that is exactly what makes combining data and sport so fascinating. Who is your pick to lift the trophy in 2026? Thanks, Dominik Kundel Aishwarya Srinivasan Arvind Narayanamurthy #FIFAWorldCup #WorldCup2026 #SportsAnalytics #DataAnalytics #DataScience #MachineLearning #Codex #Kaggle #FootballAnalytics #PredictiveAnalytics #DashboardDesign

It's only Week 1 of my Agentic AI course and I already built an app I'm genuinely excited to use for myself. 2-3 hours. A working personal finance tracker. And the most important decision I made had nothing to do with code. I started the way most people do — opened Claude Code and just asked it to build the app. It produced something. Technically impressive. But I had no way to steer it, and it wasn't what I actually wanted. Then the PM brain kicked in. Stop. Go back. Define the problem properly first. So I spent time with Claude just talking through the specs: — What accounts do I have? — What's the difference between spending money and moving it between accounts? — What's V0 vs what goes on the roadmap? — What am I not thinking about? That last question alone surfaced 10 things I hadn't considered. Once the spec was locked, the actual build prompt was straightforward. The app — fintrack — solves something that has always frustrated me: most finance apps misinterpret transfers as income or expense. fintrack automatically detects those transfers and correctly buckets the spending categories of that payment. V0 features: → Multi-account support across checking, savings, credit cards, investments → CSV import from any bank with saved format profiles → Automatic transfer matching across accounts → Custom dashboard with net worth, net in/out, time series, spending by category V1 is already specced — Same flows, real persistence. Code + PRD on GitHub 👇 lnkd.in/eKh_dRsJ Thank you to Aishwarya Srinivasan and Arvind for building this amazing course— 2 lectures, 3 guest lectures, and office hours in Week 1 alone. What an enriching start. #AgenticAI #BuildInPublic #AI #ProductManagement

A YouTube video from Aishwarya Srinivasan made me rethink a skill I've been underestimating for years. She explained Python data types using simple analogies: 𝑨 𝒍𝒊𝒔𝒕 𝒊𝒔 𝒂 𝒍𝒐𝒄𝒌𝒆𝒓 - 𝒑𝒖𝒕 𝒕𝒉𝒊𝒏𝒈𝒔 𝒊𝒏, 𝒕𝒂𝒌𝒆 𝒕𝒉𝒊𝒏𝒈𝒔 𝒐𝒖𝒕. 𝑨 𝒅𝒊𝒄𝒕𝒊𝒐𝒏𝒂𝒓𝒚 𝒊𝒔 𝒂 𝒑𝒉𝒐𝒏𝒆 𝒃𝒐𝒐𝒌 -𝒍𝒐𝒐𝒌 𝒖𝒑 𝒃𝒚 𝒏𝒂𝒎𝒆. 𝑨 𝒕𝒖𝒑𝒍𝒆 𝒊𝒔 𝒂 𝒔𝒆𝒂𝒍𝒆𝒅 𝒆𝒏𝒗𝒆𝒍𝒐𝒑𝒆 - 𝒐𝒏𝒄𝒆 𝒄𝒍𝒐𝒔𝒆𝒅, 𝒊𝒕 𝒅𝒐𝒆𝒔𝒏'𝒕 𝒄𝒉𝒂𝒏𝒈𝒆. 𝑨 𝒔𝒆𝒕 𝒊𝒔 𝒂 𝒈𝒖𝒆𝒔𝒕 𝒍𝒊𝒔𝒕 - 𝒏𝒐 𝒅𝒖𝒑𝒍𝒊𝒄𝒂𝒕𝒆𝒔 𝒂𝒍𝒍𝒐𝒘𝒆𝒅. I've spent 15 years working in data and engineering. I already knew what lists, dictionaries, tuples, and sets were. But that explanation stuck with me more than any technical definition ever had. And it made me realize something. We're entering a world where knowledge is becoming abundant. 𝘈𝘐 𝘤𝘢𝘯 𝘨𝘦𝘯𝘦𝘳𝘢𝘵𝘦 𝘤𝘰𝘥𝘦. 𝘈𝘐 𝘤𝘢𝘯 𝘦𝘹𝘱𝘭𝘢𝘪𝘯 𝘤𝘰𝘯𝘤𝘦𝘱𝘵𝘴. 𝘈𝘐 𝘤𝘢𝘯 𝘢𝘯𝘴𝘸𝘦𝘳 𝘲𝘶𝘦𝘴𝘵𝘪𝘰𝘯𝘴. The advantage is no longer having information. The advantage is making information understandable. For years, technical depth was enough. Now technical depth plus communication is becoming a superpower. The people who thrive in the AI era won't just be the ones who understand technology. They'll be the ones who can make it click for everyone else. 𝐈𝐧 𝐚 𝐰𝐨𝐫𝐥𝐝 𝐰𝐡𝐞𝐫𝐞 𝐀𝐈 𝐜𝐚𝐧 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐞 𝐢𝐧𝐟𝐨𝐫𝐦𝐚𝐭𝐢𝐨𝐧, 𝐭𝐡𝐞 𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐭𝐨 𝐦𝐚𝐤𝐞 𝐢𝐧𝐟𝐨𝐫𝐦𝐚𝐭𝐢𝐨𝐧 𝐮𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐚𝐛𝐥𝐞 𝐦𝐚𝐲 𝐛𝐞𝐜𝐨𝐦𝐞 𝐭𝐡𝐞 𝐦𝐨𝐬𝐭 𝐯𝐚𝐥𝐮𝐚𝐛𝐥𝐞 𝐬𝐤𝐢𝐥𝐥 𝐨𝐟 𝐚𝐥𝐥. #DataEngineering #Python #ArtificialIntelligence #Communication

Attended a great guest lecture by Dominik Kundel from OpenAI on how coding agents like Codex are evolving beyond autocomplete and pair programming into fully delegated agents. The biggest takeaway: getting value from agents is not just about writing perfect prompts. It is about giving them the right context, validation tools, and verification workflows. Dom shared how agents can use connected context from tools like Slack, Google Drive, Gmail, Calendar, GitHub, and docs to work more like a teammate than a simple assistant. He also emphasized the importance of linters, tests, previews, review artifacts, and feedback loops so agents can validate their own work and make human review easier. Really insightful session on how agentic workflows can support not just engineering, but also content, documentation, marketing, operations, and everyday productivity. Thank you, Dominik Kundel, for sharing practical examples and best practices. A huge thank you to Aishwarya Srinivasan and Arvind Narayanamurthy for organizing and hosting an outstanding guest lecture as part of the Mastering Agentic AI program.

Built my first AI-powered Streamlit application! 🤩 As part of Week 1 of the Mastering Agentic AI Maven course, I built an ESG Greenwashing Detector – a Streamlit application that analyses ESG disclosure data and helps identify potential greenwashing risks across companies and sectors. When I was thinking about what application to build, I wanted to connect it to something I am currently studying. Since one of my modules is Responsible International Business and Strategic Risk Management, I decided to explore the topic of ESG reporting and greenwashing. This project gave me an opportunity to combine concepts from my studies with AI-powered application development. The application includes: -ESG data upload and analysis -Greenwashing risk dashboard -Company ESG deep-dive analysis -AI-generated ESG assessments -AI-powered ESG analyst chat using Groq Coming from a non-coding background, I honestly didn't expect to build and deploy an application within the first week. This experience showed me how powerful AI-assisted development can be and how quickly ideas can be turned into functional applications. A big thank you to Aishwarya Srinivasan, Arvind Narayanamurthy and the team for the excellent training sessions, detailed documentation, and support. ❤️ This is just our Week 1 learning project, and I'm excited to continue building more in the coming weeks. 😊 Live App: lnkd.in/ew_-Cr5T GitHub Repository: lnkd.in/ewvhVWeT

A little while ago I shared the first version of a project from the Mastering AI Agents course — an interactive world map of what every country grows. Since then I've kept iterating — and that continuous building is exactly where the real learning is happening. Every new version teaches me something I couldn't have picked up any other way. What's new this time: • A cleaner, tabbed UI with summary cards and a richer, easier-to-read map • A built-in AI assistant that answers questions about the data — "Which countries produced the most wheat in 1985?" • Built with both Claude and Codex — they complement each other wonderfully • The Superpowers plugin made Codex even more capable for multi-step building • Designed to stay free and efficient: a focused data context, tool-based queries, and a smart provider fallback (Gemini → Groq → Cerebras → OpenRouter → Mistral) See how it's evolved 👇 🔹 First version: lnkd.in/euUWKc-r 🔸 Latest version: lnkd.in/erqBKhH7 💻 Code: lnkd.in/eAk46ifj Truly grateful to Aishwarya Srinivasan, Arvind Narayanamurthy, and The Gen Academy — the encouragement to keep building and sharing in public makes all the difference. Excited for what's next. 🌱 #AIAgents #VibeCoding #BuildInPublic #MasteringAIAgents

The Mastering Agentic AI bootcamp at The Gen Academy has one theme: #BuildInPublic . So here's Week 1. With my background in data and analytics I wanted to build something end-to-end, production-grade, and close to that world. Meet ToyWorld 🧸 - a full-stack Retail Management System covering everything from point-of-sale to sales analytics to IT administration, with role-based access control throughout. Built around an imaginary toy store, but designed to plug into any small retail business from day one. The architecture diagram below gives a quick sense of how it all fits together. Backed by a star schema in Supabase with 10,000+ sales rows, 500 customers, 300+ SKUs, 2 years of data with seasonality modeling, and 16 interactive Plotly charts across four analytic dimensions. Built to feel like production. 🛠️ Stack: Claude Code (+ Superpowers) · Streamlit · Pandas · Plotly · Supabase · Faker I'm excited about what AI-assisted development makes possible. Sharing the work as it happens feels like the most honest way to learn and if it sparks an idea or helps someone on a similar path, even better. 🔗 GitHub: lnkd.in/ggazramp 🚀 Live App: lnkd.in/gjXmvQjP More weeks to come. 🚀 Aishwarya Srinivasan Arvind Narayanamurthy #BuildInPublic #AgenticAI #AIAssistedCoding #ClaudeCode #Superpowers #Streamlit #Supabase #DataAndAnalytics #RetailTech #AIBootcamp #GenerativeAI #SolutionsArchitect

A few weeks ago, I shared my journey back into the Data + AI space with a structured plan. Here’s what has unfolded since then. I didn’t just study; I built lnkd.in/eNTK9Bp3. Watch Demo lnkd.in/e3whCMFz My passion for structured learning led me to enroll in the Mastering Agentic AI course by Gen Academy, which became a pivotal moment for me. The course not only deepened my understanding of agentic AI but also instilled in me the confidence to start vibe coding. This approach allowed me to build with AI rather than learn about it. Vibe coding involves describing what you want to an AI coding agent and iterating through conversation instead of writing every line yourself. My tool of choice has been Claude Code. My first real project was an Ayurveda Prakriti Assessment app, a web app designed to help users discover their dosha constitution and provide personalized daily routines, diet, exercise, and seasonal lifestyle guidance based on 5,000-year-old Ayurvedic wisdom. I transformed an idea into a deployed app, all on my own, with a clear vision and an AI pair programmer. This experience taught me that the engineers who will excel in the next wave are those who can think clearly, break down problems, and guide AI to execute. Vibe coding is not about writing less code; it’s about thinking more clearly about what you’re building and why. Here’s my current learning stack: → Vibe coding with Claude Code—building real projects, not tutorials → Mastering Agentic AI (Gen Academy)—understanding agent architectures, tool use, and reasoning patterns → Strengthening Python, SQL, and data pipeline fundamentals daily → Connecting the dots between data engineering and AI-first development The career break provided clarity, and my comeback is generating momentum. If you’re also upskilling in AI, returning after a break, or curious about vibe coding, let’s connect. I’m learning in public and sharing my journey along the And if you want to try the app I built: lnkd.in/eNTK9Bp3 github link : lnkd.in/e4AiTp7V Demo Video: lnkd.in/e3whCMFz More to come. 🚀 Aishwarya Srinivasan Arvind Narayanamurthy #VibeCoding #AgenticAI #ClaudeCode #DataEngineering #AI #CareerRestart #BuildInPublic #WomenInTech #LearnInPublic #Python #GenAcademy

Small win, nothing fancy 🎉 I deployed StartupScope AI on Hugging Face Spaces — an amateur startup idea validation tool built with LangChain, Streamlit, and OpenAI. This was a simple project to learn how Hugging Face works and how quickly I can go from an idea to a deployed application. What I have used: LangChain chains and prompts Streamlit for the frontend OpenAI for idea analysis Hugging Face Spaces for deployment There's always something satisfying about taking a project from a local machine to something that anyone can access and use. Aishwarya Srinivasan Arvind Narayanamurthy #AIEngineering #LangChain #Streamlit #OpenAI #HuggingFace

Two years ago, I moved to the United States and bought my first car. A few months in, something felt off. I took it to a mechanic. They plugged in a scanner, spent 10 minutes looking at it, and handed me a $200 diagnosis bill. Then they told me I needed a new catalytic converter, a $2,000 repair. I didn’t know enough to push back. I didn’t know what the data meant. I just had to trust them. That moment stuck with me. So for my Week 1 project at the Mastering Agentic AI Bootcamp, I didn’t build the standard stock portfolio analyzer. I built the tool I wish I had that day. The 𝗢𝗕𝗗𝟮 𝗘𝗻𝗴𝗶𝗻𝗲 𝗗𝗶𝗮𝗴𝗻𝗼𝘀𝘁𝗶𝗰𝘀 𝗧𝗲𝗹𝗲𝗺𝗲𝘁𝗿𝘆 𝗟𝗮𝗯 - a dashboard that reads the same sensor data a professional scan tool reads and explains, in plain English, what is wrong with your car and what you actually need to fix. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗱𝗼𝗲𝘀: → Detects vacuum leaks, rich/lean conditions, and fuel trim anomalies → Shows where the fault appears - idle, acceleration, or cruise → Assesses catalytic converter health from O₂ sensor waveforms → Cross-references faults with NHTSA recall and complaint data If I had had this tool two years ago, I would have known it was a vacuum leak a $30 fix, not a catalytic converter replacement. 𝗪𝗵𝗮𝘁 𝗜 𝗹𝗲𝗮𝗿𝗻𝗲𝗱 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗶𝘁: ✦ Define the problem before touching the keyboard ✦ Modular code survives feature additions ✦ src/ modules should never import Streamlit ✦ git checkout develop -- file.py is underrated ✦ If pandas downloads a .tar.gz on Streamlit Cloud, stop the build 50 commits. 3 versioned releases. 18 edge cases validated. 𝗕𝘂𝘁 𝘁𝗵𝗶𝘀 𝗶𝘀 𝗷𝘂𝘀𝘁 𝘁𝗵𝗲 𝘀𝘁𝗮𝗿𝘁. A $30 OBD2 dongle can extract dozens of live sensor readings from your car. Right now, this dashboard covers fuel trim, O₂ sensors, and RPM dynamics. But modern cars have hundreds of sensors - transmission temperature, ABS, throttle position, battery health, emissions systems. The vision -> a tool that any normal person can plug into their car, upload the data, and understand what is wrong before walking into a mechanic’s shop. No more $200 diagnostic fees. No more replacing a $2,000 catalytic converter when the real problem is a $30 vacuum hose. This tool is free. Open source. Massive thanks to Aishwarya Srinivasan and Arvind Narayanamurthy for building a AI Cohort that pushed me to build things that actually matter. Live app link and GitHub repo in the comments below!

Attended an insightful session with Harsh Vardhan Jain on “Making AI Agents Secure” by Aishwarya Srinivasan along with Arvind Narayanamurthy and Harsh Nath Jha and it genuinely changed the way I think about AI security in the agentic era. As AI agents become capable of writing code, calling APIs, accessing enterprise systems, and making autonomous decisions, traditional security models are no longer enough. The discussion around agent identity abuse, non-human identity sprawl, excessive permissions, long-lived tokens, and autonomous access to infrastructure was particularly eye-opening. One slide that stood out to me highlighted how a single compromised AI agent identity could potentially cascade into access across: GitHub repositories, Slack workspaces, AWS resources, etc. The session strongly emphasized why future AI systems need: Ephemeral credentials, Just-in-time access provisioning, Fine-grained permissions, Strong governance and observability layers for agents. As someone currently working in the Application Security space while also being deeply interested in AI systems, it was exciting to see the intersection of GenAI + Security being discussed with such practical depth. The future of cybersecurity will definitely involve securing not just humans, but autonomous AI identities as well. #CyberSecurity #AISecurity #GenAI #LLM #AgenticAI

Just wrapped up an incredibly intense session diving deep into Agentic Security and the operational reality of the Agentic SOC. The material provided a necessary framework for how we must evolve our defenses. lnkd.in/ge8k_jUU This experience perfectly illustrates the critical inflection point we are all facing: 🕰️ The Past: Traditional security stacks were built for static environments, leaving massive gaps when dealing with autonomous AI agents that execute code and act across infrastructure at machine speed. 🧭 The Present: We see that legacy controls simply fail in the Agentic era. The real work now is moving beyond reactive alerts to architecting a true Agentic Security Mesh—a self-healing framework for governing autonomous systems. 🔮 The Future: The future doesn't start with just listening; it starts with building the guardrails. This realization confirms why our focus on agent-driven governance and proactive control is so essential right now. The practical takeaway? We must move from theory to application—like taking a specific SOP and automating its security controls. This is exactly the hands-on engineering required for true resilience in this new landscape. Huge appreciation to Harsh Nath Jha and the organizers Aishwarya Srinivasan , Arvind Narayanamurthy for framing this necessary evolution. The work we are doing to build these future defenses is more vital than ever. #AgenticAI #AISecurity #DevSecOps #SecurityAutomation #FutureOfDefense

From Text to SQL: How SQL AI Agents are Transforming Data Access Wrapped up an incredible session on 𝗦𝗤𝗟 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 with Rami Krispin, where he shared his rich experience and production ready play-book. Encapsulated - Top 5 pointers (provided AI generated image for illustration only) // Context is King, Not the Model You don't need the fanciest LLM to build powerful SQL agents. A thoughtfully engineered prompt with rich context (schema details, categorical values, business logic) can outperform expensive models. It's about 𝘩𝘰𝘸 you ask, not just 𝘸𝘩𝘰 you ask. // The Agent Loop is Everything It's not just LLM → Output. Real agents iterate through reasoning → tool calls → observation → refinement. This dynamic execution path is what separates true agents from simple API wrappers. // Safety Guardrails Aren't Optional : They're Essential Read-only access, query validation, row limits, logs and error handling aren't afterthoughts. They're foundational. One misconfigured query can bring down your entire database. Period. // Memory Management Changes the Game Adding conversation history transforms single-shot queries into meaningful dialogue. Users can ask follow-ups naturally. But manage it wisely, unbounded memory explodes your token costs. // Evaluation & Fallback Strategies = Smart Cost Optimization Test your models systematically. Sometimes GPT-4 mini outperforms GPT-4 Turbo at 1/4 the cost. Build intelligent fallback mechanisms; simple models for routine queries, powerful ones for edge cases. A heartfelt thank you to Aishwarya Srinivasan, Arvind Narayanamurthy for setting up context/de-briefing about AI agent, and bringing in Rami Krispin (SME) for this Lightening session. Your generosity in sharing examples/use cases, live demos are highly appreciated. I gained additional insights and I'm sure all the participants who attended "live" would have benefitted, and reach out to discord community for Q&A/follow-up. #SQLAIAgents #AIEngineering #LLMs #DataAccess #ProductionAI #TechEducation #LinkedInLearning

Attended a session on Building SQL AI Agents at The Gen Academy today. Honest take? Parts of it felt overwhelming. As someone just starting to explore AI, I could follow the high-level ideas, but the deeper discussions around context, architecture, and guardrails made one thing clear: There’s a lot to learn here. And that was the most useful part. Before writing a single line of code, the session broke down how agents actually work: • Agents don’t respond in one step, they run a loop: Think → Act → Observe until the goal is met • There are key levers you can control to shape that loop • Every agent is built from a set of core building blocks • Not every problem needs an agent, knowing when to use one matters Understanding these fundamentals before building matters more than I expected. Then it got more practical, what it takes to move from idea to something usable: • The gap between a polished demo and a production system is significant • Context is more than a prompt, it includes instructions, knowledge, and state • Guardrails and validation are what make these systems reliable Instead of trying to learn everything upfront, I’ve decided to start building a simple SQL agent and learn through the process. Thanks to Aishwarya Srinivasan, Arvind Narayanamurthy and Rami Krispin for keeping it practical and grounded. #AIAgents #GenAI #SQLAgent #DataEngineering #DataAnalytics

Fun and learning sounds like extremities until there is passion in it. Aishwarya Srinivasan you made the evening a fun AI enlightening meetup and the truth is we genuinely learn forever. Looking forward to Cohort 2

Attended an incredible session on 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗦𝗤𝗟 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 hosted by Aishwarya Srinivasan, Arvind Narayanamurthy, and Rami Krispin. 𝗜𝗳 𝗮𝗴𝗲𝗻𝘁𝘀 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗯𝗿𝗮𝗶𝗻 𝗼𝗳 𝘁𝗵𝗲 𝗔𝗜 𝗲𝗰𝗼𝘀𝘆𝘀𝘁𝗲𝗺, 𝘁𝗵𝗲𝗻 𝗱𝗮𝘁𝗮 𝗮𝗻𝗱 𝗮𝗻𝗮𝗹𝘆𝘁𝗶𝗰𝘀 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗳𝘂𝗲𝗹. So I went ahead and built my own SQL AI Agent from scratch. What I built: ✅ Natural language to SQL using Ollama (free, no API key) ✅ SQLite database with zero setup ✅ Streamlit UI with query history and CSV export ✅ Runs 100% locally, no cloud, no cost Huge thanks to Aishwarya, Arvind, and Rami for making this so hands-on. Sessions like this push you to stop watching and start building. Check out the repo here: 🔗 lnkd.in/ebX4Dqix #AI #SQLAgent #Ollama #LLM #DataScience #ArtificialIntelligence #DataEngineering #MachineLearning #Python #Canada

Had a great time at the AI meetup hosted by The Gen Academy in SF this week. One thing from Aishwarya Srinivasan’s talk that really stayed with me was how much GTM for developer tools has changed. Feels like polished marketing and generic funnels matter way less now. People just want to see the product actually work. The teams standing out are the ones shipping fast, building in public, and earning trust from developers over time. Really enjoyed the conversations and meeting more people building in the space. #AI #DataScience #GenAI #AIAgents

Great People.. Better Ideas...⚡️ Just wrapped up an incredible day at the AI Meetup in Manhattan New York. It’s a room full of builders, founders, and enthusiasts who are actually making it happen. My 3 big takeaways from the night: Speed is the new currency: The gap between "idea" and "deployed product" is shrinking by the day. Community > Code: The most powerful breakthroughs happen when we step away from the screen and start talking to each other. NYC is the place to be: The density of talent here is unmatched. Huge thanks to the organizers Aishwarya Srinivasan Arvind Narayanamurthy Aditya Suresh and everyone I chatted with today. If we met, let’s keep the conversation going! 👇 Special thanks to Anna Mai from Pinecone for the incredible insights. #AI #NYCTech #ArtificialIntelligence #Networking #Innovation #GenerativeAI #Newyork

I spent a few hours at the Pinecone x The Gen Academy, AI Meetup in NYC - a gathering of builders, founders, and AI-curious folks! Anna Mai (Pinecone): for the in-depth presentation on the Pinecone Nexus model, a Knowledge Engine for Agents, and walking through a live demo. Seeing it in motion, not just on slides, is what made the capabilities click for me. Truly appreciated the time and effort - a special thank you! Arvind Narayanamurthy - listening to you go technical on RAG and Evals, then getting a quick chat in person, was a highlight. The way you frame AI concepts through practical, working models is something I'm excited to learn! Aditya Suresh - thank you for the friendly and generous conversation on Product Management for AI platforms, and for breaking down what it looks like to build AI infrastructure. Didn't realize how much I needed those insights until I had them. Looking forward to staying in touch and meeting you in the future! Aishwarya Srinivasan - I'm a long-time admirer of your talent and consistent hard work, and your passion for educating people about AI at every level. Your RAG Explained video is one of my favorites - it sparked my interest and motivated me to keep learning to become AI-native. Meeting you in person and hearing you teach in real time was a fan moment! The AI interactive quiz was cool & fun, and walking out with copies of Patterns for Building AI Agents and Principles of Building AI Agents (Sam Bhagwat, Mastra.ai) was the key takeaway! For me, days like this are part of a deliberate shift - evolving to be AI-native. Getting fluent in RAG, LLMs, evals, and security, and genuinely excited to integrate everything I'm learning with my Product Management principles and experience on AI platforms! #AI #ProductManagement #RAG #AINative

That moment when you finally understand what everyone's talking about? Just had that at an AI event in NYC Attended this exclusive AI meetup at Pinecone HQ, and the vibe was immaculate — no corporate fluff, just actual builders and engineers talking about what they're shipping. The conversations went beyond theory; people were sharing how they're actually using AI in their work, their strategies for handling tokens efficiently, and how they're navigating privacy and security concerns. That's the real stuff that matters. The panel with Aishwarya Srinivasan and Arvind Narayanamurthy broke down how to build production-ready AI systems that actually work in the real world. We went deep into RAG systems, agentic workflows, and the practical challenges people face daily. Watching PMs, engineers, and builders exchange ideas and real use cases was genuinely valuable. Honestly, I felt like I was lagging on the whole AI thing, but being here flipped a switch. Now I get it — it's not about chasing every update, it's about building, shipping, and getting actual visibility on what works. That's the move.

Excited to share that I’ve successfully completed the Post Graduate Program in Artificial Intelligence and Machine Learning: Business Applications from The University of Texas at Austin – McCombs School of Business. This journey gave me hands-on exposure to machine learning, deep learning, NLP, predictive analytics, Generative AI, and Retrieval-Augmented Generation (RAG) architectures, along with building and evaluating AI-driven business solutions. It was exciting to learn how AI/ML can be applied to create more intelligent, scalable, and user-focused products. I’m especially interested in the growing space of GenAI applications, RAG-based systems, and AI-powered experiences. Balancing full-time work, learning, and family responsibilities made this journey both challenging and rewarding, and I’m incredibly grateful for the support system that helped me through it. A special thank you to my manager, Shawn Hooley, and my team at work for their encouragement and support in fostering a culture of continuous learning, growth, and exploration of new technologies. The learning journey continues, and I’m excited to next explore Agentic AI, autonomous systems, and advanced GenAI workflows through NVIDIA partnered Agentic AI program by Aishwarya Srinivasan and Arvind Narayanamurthy. As a working mom, one of my biggest motivations was to set an example for my kids that learning and growth never stop. Sharing this milestone on Mother’s Day makes it even more meaningful. Happy Mother’s Day to all the moms balancing dreams, responsibilities, careers, and family every single day. #ArtificialIntelligence #MachineLearning #GenerativeAI #RAG #LLM #NLP #WomenInTech #WorkingMom #UTAustin #MothersDay

𝐅𝐫𝐨𝐦 𝐅𝐥𝐨𝐩𝐩𝐲 𝐃𝐢𝐬𝐤𝐬 𝐭𝐨 𝐌𝐮𝐥𝐭𝐢-𝐀𝐠𝐞𝐧𝐭 𝐒𝐲𝐬𝐭𝐞𝐦𝐬: 𝐓𝐡𝐢𝐬 𝐒𝐡𝐢𝐟𝐭 𝐅𝐞𝐞𝐥𝐬 𝐃𝐢𝐟𝐟𝐞𝐫𝐞𝐧𝐭. 𝖣𝖾𝖼𝖺𝖽𝖾𝗌 𝖺𝗀𝗈, 𝖨 𝖼𝖺𝗋𝗋𝗂𝖾𝖽 𝖺 𝖿𝗅𝗈𝗉𝗉𝗒 𝖽𝗂𝗌𝗄 𝖺𝖼𝗋𝗈𝗌𝗌 𝗍𝗈𝖶𝗇 𝗂𝗇 𝖺𝗇 𝖺𝗎𝖙𝗈𝗋𝗂𝖼𝗄𝗌𝗁𝖺𝗐. 𝖩𝗎𝗌𝗍 𝖳𝗈 𝗂𝗇𝗌𝗍𝖺𝗅𝗅 𝖺 𝗌𝗈𝖿𝖾 𝖺𝗋𝖾 𝖺 𝖺𝗐𝗇’𝗍 𝖺 𝖼𝖫𝗂𝖼𝗄. 𝖣𝖾𝗉𝗅𝗈𝗒𝗆𝖾𝗇𝗍 𝗐𝗮𝖾 𝖺 𝖩𝗈𝗎𝗋𝗇𝖾𝖸. 𝐓𝐡𝐞𝐧 𝐭𝐡𝐞 𝐰𝐨𝐫𝐥𝐝 𝐜𝐡𝐚𝐧𝐠𝐞𝐝. 𝖲𝖺𝖺𝖲. 𝖢𝗅𝗈𝗎𝖽. 𝖶𝗈𝗋𝗄𝖿𝗅𝗈𝗐𝗌. 𝖲𝗈𝗙𝖳𝖶𝖺𝗋𝖾 𝖻𝖾𝖼𝖺𝗆𝖾 𝖿𝖺𝗌𝗍𝖾𝗋 𝖺𝗇𝖽 𝖲𝖼𝖺𝗅𝖺𝖻𝗅𝖾. 𝐁𝐮𝐭 𝐭𝐡𝐢𝐬 𝐭𝐢𝐦𝐞, 𝐭𝐡𝐞 𝐥𝐞𝐚𝐩 𝐢𝐬 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐚𝐥. 𝖩𝗎𝗌𝗍 𝖠 𝗒𝖾𝖺𝗋 𝗈𝗋 𝗍𝖂𝗈 𝖺𝗀𝗈, 𝗐𝖾 𝖜𝖾𝗋𝖾 𝖿𝗈𝖼𝗎𝗌𝖾𝖽 𝗈𝗇 𝗂𝗇𝖽𝗂𝗏𝗂𝖽𝗎𝖺𝗅 𝗉𝗋𝗈𝗆𝗉𝗍𝗌 𝖺𝗇𝖽 𝖫𝖫𝖬 𝗈𝗎𝗍𝗉𝗎𝗍𝗌. 𝖳𝗁𝖾𝗇 𝖼𝖺𝗆𝖾 𝖜𝖾𝗋𝖽𝖿𝗅𝗈𝗐 𝗈𝗋𝗖𝗁𝖾𝗌𝗍𝗋𝖺𝗍𝖾𝗈𝗇—𝗆𝗈𝖵𝗂𝗇𝗀 𝖻𝖾𝗒𝗈𝗇𝖽 𝖼𝗁𝖺𝗍 𝗍𝗈 𝖲𝗍𝗋𝗎𝖼𝗍𝗎𝗋𝖾𝖽 𝖺𝗎𝗍𝗈𝗆𝖺𝗍𝗂𝗈𝗇. 𝖨𝖳 𝗐𝖺𝗌 𝗉𝗋𝗈𝗀𝗋𝖾𝗌𝖾𝖽, 𝖻𝗎𝗍 𝗂𝗍 𝗐𝖺𝗌 𝗋𝗂𝗀𝗂𝖽. 𝖶𝗂𝗍𝗁 𝖬𝗒 𝖻𝖺𝖼𝗄𝗀𝗋𝗈𝗎𝗇𝖽 𝗂𝗇 𝖤𝖳𝖫 𝖺𝗇𝖽 𝗉𝗋𝗈𝖼𝖾𝗌𝗌 𝗆𝖺𝗉𝗉𝗂𝗇𝗀, 𝐈 𝐜𝐨𝐮𝐥𝐝 𝐛𝐮𝐢𝐥𝐝 𝐭𝐡𝐞 𝐥𝐨𝐠𝐢𝐜. 𝖡𝖴𝗍 𝖨𝗍 𝖜𝖺𝗌 𝖲𝗍𝗂𝗅𝗅 𝖲𝗈𝗆𝖾 𝗐𝖾𝗋𝗄. 𝖤𝖵𝖾𝗋𝗒 𝖻𝗋𝖺𝗇𝖼𝖧 𝗁𝖺𝖽 𝖳𝗈 𝖻𝖾 𝖼𝗈𝖽𝖾𝖽. 𝖤𝗏𝖾𝗋𝗒 𝖾𝖽𝗀𝖾 𝖼𝖺𝗌𝖾 𝗉𝗋𝖾𝖽𝗂𝖼𝗍𝖾𝖽. 𝖨𝖳 𝗐𝖺𝗌 𝖺𝖚𝖳𝗈𝗆𝖺𝖳𝗂𝗈𝗇, 𝖻𝗎𝗍 𝗂𝗍 𝖜𝖺𝗌𝗇'𝗍 𝖸𝖾𝖹 𝗒𝖾𝖳 "𝗂𝗇𝖳𝖾𝗅𝖫𝗂𝖾𝗇𝖾𝖽". 𝐓𝐡𝐞𝐧 𝐜𝐚𝐦𝐞 𝐭𝐡𝐞 𝐬𝐡𝐢𝐟𝐭 𝐭𝐨 𝐀𝐠𝐞𝐧𝐭𝐬. 𝖨 𝗆𝗈𝗏𝖾𝖽 𝖿𝗋𝗈𝗆 𝖻𝗎𝗂𝗅𝖽𝗂𝗇𝗀 𝖲𝗍𝖺𝗍𝗂𝖼 𝖿𝗅𝗈𝗐𝗌 𝗍𝗈 𝗈𝗋𝖼𝗁𝖾𝖾𝖲𝗍𝗋𝖺𝖾𝗅𝗂𝗇𝗀 𝗌𝗒𝖲𝗍𝖾𝗆𝗌 𝗐𝗂𝖳𝗁 𝑪𝒍𝒂𝒖𝒅𝒆 𝑪𝒐𝒘𝒐𝒓𝒌 𝖺𝗇𝖽 𝑪𝒓𝒆𝒘𝑨𝑰. 𝐂𝐫𝐞𝐰𝐀𝐈, 𝗂𝗇 𝗉𝖺𝗋𝗍𝗂𝖼𝗎𝗅𝖺𝗋, 𝗂𝗍 𝖼𝗁𝖺𝗇𝖼𝖾𝖽 𝗍𝗁𝖾 𝗀𝖺𝗆𝖾 𝖿𝗈𝗋 𝗆𝖾. 𝖨𝗇𝖲𝗍𝖾𝖺𝖽 𝗈𝖿 𝖽𝖾𝖿𝗂𝗇𝗂𝗇𝗀 𝖾𝗏𝖾𝗋𝖸 𝖬𝗂𝗆𝗋𝗈-𝖲𝗍𝖾𝗉, 𝐈 𝐬𝐭𝐚𝐫𝐭𝐞𝐝 𝐝𝐞𝐬𝐜𝐫𝐢𝐛𝐢𝐧𝐠 𝐭𝐡𝐞 𝐢𝐧𝐭𝐞𝐧𝐭 𝐚𝐧𝐝 𝐫𝐨𝐥𝐞𝐬. 𝐀𝐧𝐝 𝐁𝐨𝐨𝐦 – 𝖢𝗋𝗈𝗐 𝖠𝖨 built a multi-agent system end-to-end. 𝖳𝗁𝖾 𝗌𝖾𝗌𝗍𝖾𝗆 𝖽𝖾𝗌𝗂𝖌𝗇𝗌 𝗍𝗁𝖾 𝗉𝖺𝗍𝗁. 𝖠𝗀𝖾𝗇𝖳𝗌 𝖼𝗈𝗅𝗅𝖺𝖻𝗈𝖗𝖺𝖳𝖾. 𝖤𝖷𝖾𝖼𝖼𝗎𝗍𝗂𝗈𝗇 𝗁𝖺𝗉𝗉𝖾𝗇𝗌 𝖺𝗍 𝗌𝖼𝖺𝗅𝖾. 𝖨’𝖵𝖾 𝗌𝗁𝗂𝖿𝗍𝖾𝖽 𝖿𝗋𝗈𝗆 𝖻𝗎𝗂𝗅𝖽𝖾𝖽 𝗍𝗁𝖾 𝗍𝗋𝖺𝖼𝗄𝗌 𝗍𝗈 𝐝𝐢𝐫𝐞𝐜𝐭𝐢𝐧𝐠 𝐭𝐡𝐞 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞. 𝐀𝐬 𝐚𝐧 𝐀𝐈 𝐩𝐫𝐚𝐜𝐭𝐢𝐭𝐢𝐨𝐧𝐞𝐫, 𝐈’𝐯𝐞 𝐥𝐞𝐚𝐫𝐧𝐞𝐝 𝐭𝐡𝐚𝐭 𝐭𝐡𝐞 𝐫𝐞𝐚𝐥 𝐛𝐨𝐭𝐭𝐥𝐞𝐧𝐞𝐜𝐤 𝐢𝐬𝐧'𝐭 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞—𝐢𝐭’𝐬 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲. 𝖳𝗈𝗄𝖾𝗇𝗌 𝖺𝗋𝖾 𝗍𝗁𝖾 𝗇𝖾𝗐 "𝗀𝖺𝗌 𝗆𝗂𝗅𝖾𝖺𝗀𝖾" 𝗈𝖿 𝗍𝗁𝖾 𝖾𝗇𝗍𝖾𝗋𝗉𝗋𝗂𝖲𝖾. 𝖥𝗈𝗋 𝗍𝗈𝗄𝖾𝗇 𝗈𝗉𝗍𝗂𝗆𝗂𝖿𝖺𝗍𝗂𝗈𝗇, 𝐈’𝐯𝐞 𝐛𝐮𝐢𝐥𝐭 𝐚𝐧 𝐚𝐩𝐩 𝐨𝐧 𝐁𝐮𝐛𝐛𝐥𝐞 𝐀𝐈 – 𝐧𝐨 𝐜𝐨𝐝𝐞 𝐰𝐫𝐢𝐭𝐭𝐞𝐧. 𝑾𝒆 𝒂𝒓𝒆𝒏’𝒕 𝒋𝒖𝒔𝒕 𝒖𝒔𝒊𝒏𝒈 𝒔𝒐𝒇𝒕𝒘𝒂𝒓𝒆 𝒂𝒏𝒚𝒎𝒐𝒓𝒆. 𝑾𝒆 𝒂𝒓𝒆 𝒄𝒐-𝒄𝒓𝒆𝒂𝒕𝒊𝒏𝒈 𝒊𝒕. #AI #GenerativeAI #CrewAI #Claude #BubbleAI #RevenueOps #AppliedAI #AIOrchestration 𝐍𝐨𝐭𝐞: 𝘐𝘧 𝘺𝘰𝘶 𝘸𝘢𝘯𝘵 𝘵𝘰 𝘭𝘦𝘢𝘳𝘯 𝘊𝘭𝘢𝘶𝘥𝘦 𝘊𝘰𝘸𝘰𝘳𝘬, 𝘐 𝘩𝘪𝘨𝘩𝘭𝘺 𝘳𝘦𝘤𝘰𝘮𝘮𝘦𝘯𝘥 𝘸𝘢𝘵𝘤𝘩𝘪𝘯𝘨 𝘵𝘩𝘪𝘴 𝘵𝘶𝘵𝘰𝘳𝘪𝘢𝘭 𝘣𝘺 Aishwarya Srinivasan - 𝐥𝐢𝐧𝐤 𝐢𝐧 𝐜𝐨𝐦𝐦𝐞𝐧𝐭 𝐛𝐞𝐥𝐨𝐰!

Multimodal AI is moving us beyond text-based interactions. This article from The Gen Academy highlights how AI is rapidly evolving to interpret and respond across speech, images, and video, not just written input: lnkd.in/gE9GD_AW In practice, I’m already seeing this shift, using the speech option in ChatGPT consistently yields more context-rich, natural responses compared to text alone. This isn’t incremental progress. It’s a change in how humans and systems will interact - more intuitive, more contextual, and closer to real-world communication. Worth paying attention to, as we design the next generation of products and experiences. Thank you Aishwarya Srinivasan for sharing your expertise to the rest of the world, really appreciate your work!