work
Projects
Tools, systems, and experiments that earn their keep.
The work here is real; the polished public case studies are not done yet.
For launch, this page is a field guide to the projects I spend time on. Deeper write-ups will follow once they are genuinely useful.
-
LLMs in production need structured work boundaries — unbounded sessions drift and can't be replayed or evaluated. A CLI-first, event-sourced execution runtime constraining LLM work into typed stages (triage → deliberation → DAG planning → worker/judge execution → artifact snapshot), backed by PostgreSQL with advisory-lock append safety. TypeScript monorepo across 10+ packages, deterministic in-memory test harness, and 179 passing tests. One layer of a three-layer architecture communicating through typed contracts, not internal bindings.
Active development · private.
-
Document review in most tools is passive — opening a file leaves no durable record. A version-aware reading workspace where review is an explicit lifecycle: annotations, highlights, voice segments, and transcripts captured as a structured, append-only record. React + Vite frontend, Electron shell for IPC-backed filesystem access, OpenAI integration for narration and transcription, slash-command navigation, and Markdown export. Runs in browser, Docker, and Electron modes.
Active · private.
-
AI configurations drift across tools, producing inconsistent agent behavior when moving between Copilot, OpenCode, or Gemini CLI. A portable runtime package spanning 3 AI platforms — reusable instruction sets, durable skills, tool adapters, design doctrine, and machine-safe configuration boundaries — with non-destructive startup self-checks to detect drift without mutating local state.
Active · private.
-
Repeated raw search across a personal knowledge vault is slow at agent speed. Two complementary tools: a local KB MCP server with SQLite-backed full-text indexing and structured retrieval commands over a multi-hundred-note Obsidian vault; and a Go-based journal MCP/CLI for daily entry management — giving agents a fast, reusable retrieval surface without raw file access.
Active · private.
-
A family token economy app: per-user interest calculations, atomic Firestore transactions for award and spend operations, birthday detection with automatic bonuses, and a full command surface via Commander.js. Ran in production for two years as a real household tool. Primary vehicle for learning real-time database design and Firestore transaction patterns.
Retired · private.