Four things in active development. Three more in the lab. If you want to talk about any of them — [email protected].
featured
Playwright agent that walks a site, scores the UX, and emits a multi-page PDF. Tiered model architecture (Haiku scout → Sonnet executor → Opus advisor) with a separate eval harness that quantified which variant actually wins — the cheaper configuration, 2W/1L.
Automated UX quality audit agent built on Playwright + multi-model scoring. Cut running cost 84% by stripping images from conversation history and shipping screenshots at JPEG-40.
Private SvelteKit dashboard on Fly.io, reading a SQLite mirror of an Obsidian vault via Litestream. Modules for projects, daily notes, biometrics, and system health.
Five explicit LLMOps decisions in a working RAG pipeline: H2 structural chunking over fixed-token windows, frontmatter separation from embedded text, asymmetric query/document input types, SQLite as a queryable observability layer, and incremental indexing via content hashing.
also building
Built because sustained attention to maintenance is exactly what ADHD is worst at — so the checker had to be the thing that pays attention. Self-discovers every automation: launchd jobs, git hooks, OAuth tokens, vault integrity. Reports what to run when something breaks.
Specialist Claude agents for vault queries, hiring analysis, curriculum review, and automation design — each with typed inputs and narrow scope. An eval harness runs periodic trajectory checks to surface drift before it compounds. The architecture is LLMOps patterns made operational.
Self-paced path through the operational layer of LLM systems — observability, evals, cost routing, model gateways. Each phase maps to work already running in the portfolio: reasonable-ux, QA Agent, and Agent Panel.