The Engineering Behind Enterprise AI

  • Self-Scaffolding LLMs: How Ornith-1.0 Rewrites Its Own Harness Mid-Training

    A technical breakdown of DeepReinforce’s self-improving agentic coding models   There’s a quiet assumption baked into most RL post-training pipelines for coding agents: a human designs the harness, and the model just gets better at using it. The scaffold — memory management, error handling, tool orchestration, retry logic — stays fixed. The policy is the…

    Amit Patriwala

  • Why a 7B Parameter Model Won’t Run Comfortably on a 14 GB GPU (And Why Most Engineers Get This Wrong)

    If you’ve recently started working with Large Language Models (LLMs), you’ve probably seen a calculation like this:   7 Billion Parameters × 2 Bytes (FP16) ≈ 14 GB   At first glance, it seems perfectly reasonable to conclude: “A GPU with 14 GB of VRAM should be enough.”   Unfortunately, that’s one of the most…

    Amit Patriwala

  • Introducing PersonaGraph: The Problem Every AI Power User Ignores

    moved to https://aitechpartner.com/introducing-personagraph-the-problem-every-ai-power-user-ignores/

    Amit Patriwala

  • RAG Evaluation Techniques — A Field Guide for Multi-Turn Voice Agents

    RAG Evaluation Techniques — A Field Guide for Multi-Turn Voice Agents   How to rigorously evaluate retrieval-augmented generation (RAG) in conversational voice agents — distinct from simple document Q&A interfaces. Written from real-world implementation experience with production voice systems.   All examples use the fictional company FleetPulse (fleet telematics provider) to illustrate concepts without referencing…

    Amit Patriwala

  • Why RAG Systems Sometimes Answer Questions Nobody Asked

    A Production Lesson Every AI Engineer Eventually Learns   One of the most surprising moments when deploying a Retrieval-Augmented Generation (RAG) system to production is watching users become frustrated by an AI that appears highly intelligent but somehow feels socially unaware.   The user says: Thank You The AI responds: According to the my knowledge…

    Amit Patriwala

  • Loops: The Quiet Skill Behind Every AI System That Actually Scales

    Why the Future of AI Isn’t About Better Models—It’s About Better Loops   Every week a new AI model arrives. A larger context window. A better benchmark score. A more impressive demo.   The industry conversation usually follows the same pattern: Is GPT-5 better than Claude? Is Claude better than Gemini? Is Gemini better than…

    Amit Patriwala

  • Learn AI for Free: 10 Platforms From OpenAI, Google, Microsoft, NVIDIA & More

      A few years ago, learning Artificial Intelligence felt expensive.   People spent thousands of dollars on bootcamps, certifications, and online programs hoping to gain AI skills that could improve their careers.   Today, something remarkable has happened.   The companies building the world’s most advanced AI systems are teaching people for free.   Not…

    Amit Patriwala

  • The Hidden Context Window Problem in RAG Systems: A Real Production Incident with vLLM and Qwen3

    When Your 32K Context LLM Fails at 4K Tokens: A Production vLLM Troubleshooting Guide   One of the most common misconceptions in Generative AI systems is: “The model supports 32K context, so my application automatically supports 32K context.”   In production, that assumption can lead to unexpected failures.   Recently, we encountered a production issue…

    Amit Patriwala

  • Want to Learn AI Without Spending Thousands on Courses?

      Microsoft has made its entire AI for Beginners curriculum available for FREE on GitHub.   📚 12 Weeks | 24 Lessons | Hands-On Labs | Open Source   This isn’t a marketing tutorial or a collection of random videos. It’s a structured learning path created by Microsoft Cloud Advocates that covers the fundamentals of…

    Amit Patriwala

  • Google Gemma 4 12B: The Model That Signals a Bigger Shift in AI Infrastructure

    Over the past year, I’ve spent a considerable amount of time working with both local and production AI environments.   On one side, I’ve been experimenting with local LLMs using Ollama, testing quantized models, and exploring how much intelligence can realistically run on developer laptops.   On the other side, I’ve been deploying production workloads…

    Amit Patriwala