>Projects

>Claude OS

Turned Claude Code into a stateful personal operating system. Running it to manage this job search right now.

• ›Git-tracked markdown files ARE the memory -- TODAY.md and MEMORY.md load into every session via startup hooks. Inspectable, editable, versioned.
• ›FastMCP server with 15+ tools: calendar events, contact lookup, email triage, scheduling, analytics -- all through conversation.
• ›Multi-instance coordination: Chief orchestrates all day, Specialists spawn for deep work, Subagents run background tasks in parallel.
• ›Running it to manage this actual job search right now -- scheduling interviews, processing emails, tracking pipeline. Not a demo.

>Contoural

AI citation analysis pipeline for an information governance firm. Clients include 30% of the Fortune 500.

• ›40 hours → 10 hours per engagement (75% reduction), enabling a shift from hourly billing to fixed-price contracts.
• ›340,000+ legal citations classified at 80–90% precision through a 3-stage extraction + classification + correction pipeline.
• ›Corrections that persist: a human-in-the-loop feedback system where reviewer corrections retrain the classifier via replay -- accuracy compounds over time.
• ›Multi-LLM orchestration via LiteLLM with Instructor for structured extraction; LLM-as-judge error taxonomy for systematic quality tracking.

>Texas Hold LLM

Full-stack poker simulation where LLM agents play Texas Hold'em. Why poker is a great LLM benchmark.

• ›Three-prompt architecture: game state (cards, pot, position), conversation history (table talk), and player profile (archetype, history) unified into one context.
• ›Table talk that actually influences decisions -- agents read bluffs, make threats, adjust strategy based on what others say.
• ›8 distinct agent archetypes with personalities (Aggressive, Passive, Bluffer, etc.) and ELO rating system with K-factor calibration.
• ›Event-sourced game engine with 13 event types and transparent reasoning panels -- every agent decision is fully inspectable.

>Open Source Legislation

All 50 US states' statutes in a single queryable format with semantic search. Open sourced when the startup shut down.

• ›50 states standardized into a single schema: 2M+ legislative sections indexed with consistent hierarchy, citations, and metadata.
• ›4 embedding types per section -- text, summary, HyDE (hypothetical document embeddings), name -- for different retrieval strategies.
• ›Built pgvector semantic search for legal text before RAG became mainstream. Still receiving commits through June 2025.
• ›Open sourced when Recodify shut down. The startup is gone; the infrastructure persists as the most complete open dataset of US state law.