Maia is a chess engine trained on millions of human games to predict the exact move a human would make, rather than playing perfectly like traditional engines. When I joined the Computational Social Science Lab, it only existed as research papers and a set of Lichess bots. My role has been to turn Maia into a usable platform where people can play against the model, study their mistakes, drill openings, and analyze their own games through a human-centered lens.

I built the platform’s architecture from the ground up using TypeScript and React (Next.js), handling real-time gameplay, move prediction, mistake classification, user accounts, and interactive training tools. On the machine learning side, I helped move Maia’s inference pipeline entirely into the browser. I converted the PyTorch weights into ONNX format and ran them using the ONNX web runtime, so model inference happens locally on users’ machines with sub-second latency. This removed server costs entirely and allowed millions of games to run without dedicated GPU infrastructure.

The result is a research-grade system that functions at public scale. Maia is now the most-played chess AI on Lichess, with more than four million games, and the platform continues to serve researchers, streamers, and everyday players who want an AI that feels human rather than unbeatable.