Writing
Thoughts on building, AI, design, and the founder journey.
I built a faithfulness evaluator for a real RAG problem Build 01 of paper-to-production: RAGAS + FActScore + MVVP as a working faithfulness evaluator — 102 labelled cases, LLM κ, cost/latency, and when I would not treat the score as a gate.
→ Building Llama 3.2 From Scratch (How Modern LLMs Improved on GPT-2) Week 2 of the model-atlas series: rebuild a Llama-style decoder block in PyTorch and see why RoPE, RMSNorm, GQA, and SwiGLU became the modern default.
→ Building GPT-2 From Scratch (and Loading Real Weights) Week 1 of a 24-model series: implement every layer in PyTorch, load OpenAI's checkpoint, and see why today's LLMs are still this architecture.
→ Seven Hidden Faults in Every Tamil NLP Pipeline Unicode fragmentation, mojibake, agglutination explosions, and the romanized web your model never saw - a field audit of what goes wrong before training.
→ Building a Tiny Tamil GPT From Scratch What I learned training a decoder-only Transformer on my own data
→ The Anatomy of an RL Environment — How AI Agents Actually Learn to Write Better Code Most people think training an AI agent is about feeding it data and hoping it gets smarter. It's not. Here's what a real RL environment looks like under the hood.
→ Building FounderOS - Agents SSharing the journey is part of the process — here's why I decided to document everything I build.
→ Why I'm building in public Sharing the journey is part of the process — here's why I decided to document everything I build.
→ Notes on product thinking The mental models I keep coming back to when building products people actually want.
→