I build reinforcement-learning systems for games, mostly Magic: The Gathering, and run small experiments on AI oversight.
mtg-kernel · Rust, Python
A deterministic Magic: The Gathering rules engine in Rust with a Python self-play training and evaluation stack. Runs are reproducible, and an XMage bridge checks both rules behavior and agent strength against an independent engine.
spellbench · Python
An open protocol and tournament arena where MTG engines and bots built by different people play each other. Seat-swapped game pairs, Elo ratings with bootstrap intervals, and every rating re-derivable from the match ledger.
Selvarath Debate · Python
Manifund-funded tests of debate as AI oversight. In a pre-registered 22,140-cell run, giving a weaker judge two oracle checks raised its error by 3.9 points (95% CI 2.2 to 5.7). My part: the pilot re-analysis, a code audit that found two data-corrupting bugs, and the follow-up harness. Pilot write-up
- XMage RL agent: transformer policy trained with PPO and population-based self-play inside the XMage engine, on a Slurm cluster (one GPU head feeding 640 CPU game runners). Predecessor to mtg-kernel.
- Diffusion medical augmentation: a class-conditional DDPM that synthesizes rare skin-lesion classes (HAM10000) as training data for rare-class classification. Group project.
- Dota Science: predicting Dota 2 match outcomes from OpenDota data, 85.7% accuracy. Write-up