Know when to let the model cook
Taste is all that matters.
A collection of notes, essays, and technical deep dives by Ivan Leo.
Currently building general agents for knowledge work at Manus.
A comprehensive series on building AI coding agents from scratch, covering everything from basic tool integration to advanced features like agentic search and subagents.
Working through Stanford's CS336 and documenting what I learn at each step — from tokenization to training.
Taste is all that matters.
The messy, fast-changing path that took me from front-end engineering to LLMs.
Creating our first RL Pipeline with Verifiers
Lessons learnt from going from zero to $100M ARR in 8 months
Building a Byte-Pair Encoding tokenizer from scratch
Getting started with reinforcement learning
Models can't do much without the right context, agentic search does just that
Switch between models with your own custom router