Bailiwick Languages is a learning engine that handles placement, grading, mastery, unlocks, and the evidence families and teachers see. Those state changes live in TypeScript and Postgres behind Supabase RLS, with separate authorization paths for learners, linked family accounts, and teachers.

Misty runs my own fine-tuned 0.6b model entirely in the browser, in a Web Worker through Transformers.js and ONNX Runtime Web, with WebGPU when the device supports it. The difficult part isn't just generating text: the app has to manage model installation and caching, bounded lesson context, privacy and answer checks, rejected-output handling, and very different device limits. Misty can explain; it sits outside the code that changes grades or learner state.

I've been contributing to Inspect AI:

#4783 added serializable ModelRole(..., required=True) support and log round-trip tests. Merged.

#4900 uses Inspect's OpenAI retry classifier while polling responses.retrieve(), with asyncio and Trio coverage. Open.

Vulnerability research is where I publish application-security work.