I build stuff to learn and to actually use.
- 🏃 training-os — my own training platform with a coach that actually sees my sessions. FastAPI backend, MCP-based LLM chat that can actually query my Strava data, reverse eng of Corros with TL/ATL/CTL/ACWR analytics. Born out of frustration of lack of integration between chatbots and physiologic data.
- ♟️ chessformer — a transformer trained from scratch to play a human moves, not the best moves. Eventual goals: feeling bar instead of eval bar, cheating detection, player style imitation, clock-situation aware engine, beatable bots at every level. Currently 50M parameter model trained on 100M tokens (100k games, 6 epochs).
- 🧠 ReWiSe — my MSc research at Mila / LAMA-WeST. Knowledge extraction from LLMs with relation-wise self-consistency. Won LM-KBC 2025.
- 🎮 ConnectZero — self-play dynamics for RL agents on Connect 4.
