I build backend systems for AI tools: durable execution, failure recovery, deterministic controls, and evaluation. I also maintain software that works offline.
| Project | What to inspect |
|---|---|
| tool-journal | A TypeScript/SQLite execution journal with lease fencing, bounded retry admission, and explicit uncertainty. Start with the failure contract, run the LangGraph recovery example, or inspect the runtime experiment. |
| WordAI Community | Offline vocabulary learning in Flutter and SQLite. Inspect meaning-level progress, content-bound answer transactions, and transactional recovery and system file export. |
A failure you can reproduce: a service commits an action, then the HTTP receipt is lost. In tool-journal's synthetic HTTP example, recovery with downstream idempotency makes two calls for one effect; subsequent replay needs no service call. Process tests and experiment controls separate the journal's contribution from the provider's deduplication guarantee.
Provider keys can expire. A fixed retry admission window prevents new retry leases after the cutoff, while preserving a live owner's ability to save its receipt. The regressions also demonstrate the limit: local admission cannot stop a delayed request from arriving after the provider forgets its key.
I build TraderBear, a paper-first AI trading and research product at Awesome Bears. Its production implementation is private. tool-journal is an independent reference implementation informed by that work, with synthetic examples and its own tests; it is not a mirror of the product.
I care about evidence that someone else can check: runnable examples, explicit failure boundaries, regression tests, and design decisions with their tradeoffs.