I build machine-learning systems: reinforcement learning for games, and more recently AI-safety experiments.

A reinforcement-learning agent that plays full games of Magic: The Gathering through the XMage engine rather than a simplified environment. Transformer policy trained with PPO, self-play with an adaptive curriculum, and population-based training across a pool of Pauper decks, scaled out on a Slurm HPC cluster (one GPU head node feeding 640 parallel game runners on CPU satellites). The long-run benchmark is reliably executing a multi-step combo line, a hard credit-assignment problem. In progress since mid-2024; the repo README covers the architecture, progress history, and how to run it.

Testing failure modes of debate-style AI oversight: an honest and a dishonest LLM debater argue before a weaker judge that can spend a limited budget of oracle verification calls. The pilot found that a small oracle budget made the judge perform worse than none at all. Funded by a Manifund grant; pilot write-up on LessWrong. My contributions: the pilot re-analysis, a code audit that found two data-corrupting bugs in the oracle channel, and the re-judge harness now running the validation experiments.

Predicting Dota 2 match outcomes from OpenDota API data: exploratory analysis, regression, and a Random Forest classifier reaching 85.7% accuracy. Live write-up.