Engineering
Things I have built.
spotlit
A 6.7 MB open-weight model that ranks a web page's elements against an AI browser agent's goal, in the browser. On Mind2Web's test splits the clicked element is in its top 50 on 76-81% of steps.
rankscope
Retrieval evaluation that says how sure it is: confidence intervals on every hit rate, per-query-class breakdowns, miss attribution, rank fusion and a calibrated abstain gate for search and RAG pipelines. Pure Python with no dependencies.
agent_mux
Runs MCP servers and agent loops in production from one OpenResty/LuaJIT process: supervised stdio MCP servers, a multi-turn agent loop with streaming responses, Redis-backed sessions and budgets, auth, rate limits and metrics.
Lumen
A local-first knowledge base for AI coding agents. Sources are compiled into a SQLite knowledge graph that Claude Code, Cursor and Codex query over MCP, with BM25, TF-IDF and optional vector search fused by rank, and end-to-end encrypted sync.
FleetLM
LLM inference served by a fleet of everyday laptops behind one OpenAI-compatible API. Each node holds a whole model and dials out over one WebSocket; work goes out as leased, idempotent units, so killing a node mid-batch only retries the units it held.