I build systems that hold up — databases, platforms, and lately open infrastructure for training AI agents.
Currently building Crucible — turn any real software (a database, a CLI, a codebase, an API) into a trainable, gradable, replayable RL environment. The reward is your real software; every episode replays byte-for-byte, so every reward is auditable. MIT.
→ A 0.5B model went 5%→100% on a real SQL task with a Crucible environment as the only reward — then generalized across shell, code, and database agents. The proof.
How I build: one source of truth, reuse over rebuild. Tests ship with the change (100% coverage, CI gates). Docs move in the same commit as the code. One production path — no MVP-vs-real branches.