AI undergraduate focused on reinforcement learning. I build small, verifiable infrastructure for AI agents.
Building
- proofline — verifiable run bundles for AI and agent executions: tamper-evident evidence, first-class redaction, semantic diffs, replay, and Ed25519 signatures. Python reference implementation of an open format.
- pi-proofline — records every pi agent run as a proofline bundle. A single dependency-free TypeScript file, byte-compatible with the Python implementation and held to it by fuzzed cross-language tests in CI.
Notes & research
- BreakRL — ablation-driven reinforcement learning fundamentals in Chinese: notes, derivations, and runnable notebooks.
- clean-first-engineering — the engineering discipline I hold my agents (and myself) to, packaged as an installable agent skill.
Evidence over vibes: if an agent did something, there should be a portable record anyone can verify.