Wall of love for The Gen Academy

I spent 3 months building a multi-agent system that does 2-4 hours of AE work in ~4 minutes.   4 LangGraph agents. Cited sales strategy. Production-grade.   3 things production taught me:   -> NAIVE RAG FAILS HARD "GNC toolboxes" returned zero semantic similarity to "Sensor Fusion and Tracking Toolbox." Vector-only search = wrong products in front of customers.   Fixed it with Hybrid RAG (thanks Aishwarya Srinivasan): BM25 + vector + cross-encoder reranking. Retrieval quality is the ceiling—everything downstream inherits its failures.   -> YOU CAN'T TUNE WITHOUT EVALS Like adjusting PID gains without feedback means you're flying blind.   Built Claude-as-judge with 4 metrics + 9 deterministic checks. Went from 3.0 to 4.2 quality score across 3 iterations.   -> LLM COST IS A ROUTING PROBLEM Ollama ($0) → Groq 8B → Groq 70B dispatched by task complexity. Prompt caching cuts 90% of token costs. MIT licensed. If any of this saves you a debugging session, fork it, build on it. Repo Link in comments.   #AgenticAI #LangGraph #RAG

Most AI answers 𝘴𝘰𝘶𝘯𝘥 𝘳𝘪𝘨𝘩𝘵… but are they actually correct? One of the challenges with LLMs becomes very clear in enterprise settings. LLMs like ChatGPT, Gemini, or Claude are trained on 𝗶𝗻𝘁𝗲𝗿𝗻𝗲𝘁 𝗱𝗮𝘁𝗮, not your company’s internal policies, processes, or documents. So when someone asks: “𝘊𝘢𝘯 𝘐 𝘴𝘩𝘢𝘳𝘦 𝘢 𝘤𝘭𝘪𝘦𝘯𝘵 𝘧𝘪𝘭𝘦 𝘸𝘪𝘵𝘩 𝘢 𝘷𝘦𝘯𝘥𝘰𝘳?” The response may sound confident and logical… but it could still be 𝗺𝗶𝘀𝗮𝗹𝗶𝗴𝗻𝗲𝗱 𝘄𝗶𝘁𝗵 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗽𝗼𝗹𝗶𝗰𝘆. 𝗧𝗵𝗶𝘀 𝗶𝘀 𝘄𝗵𝗲𝗿𝗲 𝗥𝗔𝗚 (𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 𝗔𝘂𝗴𝗺𝗲𝗻𝘁𝗲𝗱 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻) 𝗰𝗼𝗺𝗲𝘀 𝗶𝗻 RAG connects LLMs to 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝘁𝗵𝗲𝘆 𝘄𝗲𝗿𝗲 𝗻𝗲𝘃𝗲𝗿 𝘁𝗿𝗮𝗶𝗻𝗲𝗱 𝗼𝗻. Instead of relying only on pre-trained knowledge, it:  1. Breaks documents into 𝗰𝗵𝘂𝗻𝗸𝘀  2. Converts them into 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀 (numerical representations of text)  3. Stores them in a 𝘃𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲  4. Uses 𝗰𝗼𝘀𝗶𝗻𝗲 𝘀𝗶𝗺𝗶𝗹𝗮𝗿𝗶𝘁𝘆(=0.97) to retrieve the most relevant context  5. Feeds that context to the LLM to generate grounded responses 𝗖𝗵𝘂𝗻𝗸𝗶𝗻𝗴 𝗶𝘀 𝗺𝗼𝗿𝗲 𝗶𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝘁 𝘁𝗵𝗮𝗻 𝗶𝘁 𝗹𝗼𝗼𝗸𝘀 The way you split data directly impacts the quality of retrieval. Common chunking strategies include:  • Fixed-length chunking  • Sentence-based chunking  • Paragraph-based chunking  • Sliding window chunking  • Semantic chunking  • Recursive chunking Better chunking → Better retrieval → Better answers And more importantly: 𝗕𝗲𝘁𝘁𝗲𝗿 𝗮𝗻𝘀𝘄𝗲𝗿𝘀 𝘄𝗶𝘁𝗵 𝗰𝗶𝘁𝗮𝘁𝗶𝗼𝗻𝘀 Which means the response is not just helpful, but 𝘁𝗿𝗮𝗰𝗲𝗮𝗯𝗹𝗲:  • You can see where the answer came from  • You can refer back to the exact company document  • The answer has real grounding, not just confidence That’s what gives AI 𝗰𝗿𝗲𝗱𝗶𝗯𝗶𝗹𝗶𝘁𝘆 𝗶𝗻 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝘂𝘀𝗲. ⚖️𝗥𝗔𝗚 𝘃𝘀 𝗙𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 𝗙𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 → Good for behavior, structure, and low latency 𝗥𝗔𝗚 → Better for dynamic, real-time, and private enterprise data For most enterprise use cases, RAG is the most effective starting point for improving LLM responses. 𝗠𝘆 𝘁𝗮𝗸𝗲𝗮𝘄𝗮𝘆 AI is not just about generating answers. It’s about:  • Retrieving the right context  • Grounding responses in real data  • Making answers 𝘃𝗲𝗿𝗶𝗳𝗶𝗮𝗯𝗹𝗲, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝗰𝗼𝗻𝘃𝗶𝗻𝗰𝗶𝗻𝗴 A quick shoutout to Arvind Narayanamurthy & Aishwarya Srinivasan from The Gen Academy for this weekend learning.

I ran 7 RAG architectures side by side on the same questions to see what actually matters. The setup comes from a deep dive by a friend of mine, Vidhyakshaya Kannan, on The Gen Academy which breaks down these RAG architectures and their tradeoffs. A few things that stood out: - Most architectures land in a similar accuracy range - Agentic RAG performs best, but costs ~4.8× more per correct answer - Self-RAG underperforms Naive RAG. More complexity, worse results It is easy to over-engineer RAG systems. Running them side by side makes that tradeoff very clear. I’ll drop the blog post and code in the comments. How are you deciding between RAG architectures in practice? Optimizing for accuracy, cost, or something else? #MachineLearning #ArtificialIntelligence #DellTech #DellProPrecision #NVIDIA

Most AI PMs are managing probabilistic systems with deterministic tools. That's why AI products quietly rot after launch. In traditional software: a button either works or it doesn't. Bugs are reproducible. If it passes staging, it passes prod. In AI: the "same" input produces different outputs. A prompt that works on 10 examples fails on the 11th. Features don't ship as "works / doesn't work", they ship with a distribution of quality. And most PMs have no idea how to measure that distribution. I mapped out the full AI PM tooling stack for 2026, the 6 layers every AI PM needs to know, the category leaders in each, and the judgment calls that actually matter: → Evaluation (where quality is defined) → Observability (seeing what actually shipped) → Prompt management (governing changes safely) → Prototyping (Bolt, Lovable, v0, Cursor) → Experimentation (shipping to users safely) → Customer feedback (keeping evals honest) The biggest mistake AI PMs make? Adopting an eval tool and never building a real dataset. Treating the tool as the deliverable instead of the scaffolding. Inspired by the Lightning Session on The AI PM Playbook by The Gen Academy, hosted by Aishwarya Srinivasan and Arvind Narayanamurthy, with guest speaker Shyamala Prayaga. Full breakdown on Substack, link in comments 👇

Yesterday, I attended the AI PM Playbook by The Gen Academy, featuring Shyamala Prayaga (Senior AI Product Manager at NVIDIA). You might wonder, what’s a software engineer doing in an AI PM session? Because I’ve seen how often we jump straight into building… and only later realize we should’ve asked better questions first. Working closely with PMs, I’ve started valuing that thought process more, how ideas are shaped before they become features. That’s what pulled me into this session. One line from the session that really stayed with me: 👉 Don’t fall for the “shiny object syndrome” in AI. It’s tempting to add AI everywhere, but the real focus should be on: • Solving the right problem (not just using AI for the sake of it) • Defining metrics & evals upfront - “build evals before building features” • Thinking about adoption - token economics, context, and real user value A good reminder that building AI products, at any level, needs clarity, intent, and responsibility, not just hype. #AI #ProductManagement #SoftwareEngineering #GenAI #Learning

AI products don’t fail because of bad models. They fail because of bad decisions. Over the weekend, I attended the Lightning Lesson on The AI PM Playbook, hosted by Aishwarya Srinivasan and Arvind Narayanamurthy (Maven × The Gen Academy), with Shyamala Prayaga (NVIDIA). It exposed a gap in how I was thinking about AI systems. I was focused on capabilities. This shifted me toward accountability. The MAP Framework (Model → Augment → Program) stood out—not just as a technique, but as a decision filter: 1. When should AI generate? 2. When should it assist? 3. And when should it stay out entirely? This matters because AI is non-deterministic. Treat it like traditional software, and you end up with "silent failures"—systems that appear correct until they break user trust. The shift for me: As someone with a technical background, I used to ask: “What can this model do?” Now, as an aspiring AI PM, I ask: “What should the system be responsible for?” I’m now applying this product-first mindset to my own projects—moving beyond demos toward reliable systems at scale. For AI builders: what’s one decision that becomes harder when moving from demo to production? #ArtificialIntelligence #NVIDIA #AIProductManagement #GenerativeAI #SystemDesign #BuildInPublic #AIEngineering #MachineLearning #ProductStrategy #TheGenAcademy