15 March 2024
Building an unbeatable tic-tac-toe agent
What tabular Q-learning teaches you about reward shaping when the state space is small enough to reason about by hand.
Reinforcement learning · Python · Game theory
Writing
Notes from the middle of the problem
28 February 2024
Democratising legal aid with retrieval
The architecture behind a RAG pipeline over Pakistani traffic law — and why the retriever, not the model, decides whether it's useful.
RAG · LLM · Legal tech
10 January 2024
Fine-tuning FLAN-T5 for a narrow task
A walkthrough of instruction fine-tuning a small model, and the point at which prompting stopped being enough.
Fine-tuning · Hugging Face