Formal methods, grammar induction, and neuro-symbolic computation — systems that show their reasoning instead of asserting it.
Lead Scientist, Data & ML at Vairified Corp. I maintain OpenSkill, and it runs in production under the platform I build.
Bayesian rating library for asymmetric multi-team, multiplayer matches, implementing the Weng–Lin models. Pinned as a dependency in Ray/RLlib and the rating layer in Neural MMO — both checkable in those projects' manifests. Reported in production at OpenAI, MultiVersus, Hunt: Showdown, and BeyondAllReason.
Two years across eighteen solver architectures in nine families, converging on one negative result about where the difficulty in ARC actually lives.
Grammar induction, monograph in progress. The line of work runs back to OpenGrammar in 2022. Open quantities: k*, the interaction order, and the renormalizability conjectures.