Engineering
Things I have built.
spotlit – element ranker for AI browser agents
Ranks a web page's elements against an agent's goal, right in the browser. The target element lands in its top 50 on 76-81% of Mind2Web test steps.
rankscope – retrieval and RAG ranking evaluation
Measures search and RAG retrieval with confidence intervals, per-query breakdowns and miss attribution. Pure Python, no dependencies.
agent_mux – production gateway for MCP servers
Runs MCP servers and multi-turn agent loops from one OpenResty process, with sessions, budgets, auth, rate limits and metrics.
Lumen – persistent memory layer for AI agents
Compiles your sources into a local SQLite knowledge graph that Claude Code, Cursor and Codex query over MCP, with hybrid search and encrypted sync.
FleetLM – distributed inference on consumer hardware
Serves LLM inference from a fleet of everyday laptops behind one OpenAI-compatible API, retrying only the work a failed node held.