I build and test AI systems, and I publish what I find, including when the result is negative.
Longevity research (computational)
- senescence-connectivity-pilot: a drug connectivity-mapping screen for cellular senescence that failed its own validation. The code, the failed tests and a withdrawn result are all public, so others can skip the same dead end.
Voice agents
- voice-agent-disclosure-testing: how a production voice agent ended up making false claims about its own configuration, and the listening-based test that now catches it.
- livekit-voice-agent: an appointment-booking agent on LiveKit Agents, with a tool runtime that keeps calls to a calendar or CRM validated, idempotent and bounded.
Every "verified" claim in these repos points to a command or a timestamped log. Where I got something wrong, the correction is dated and the original stays visible.