Projects

Hypergraph Code Explorer (HCE)

A tool that transforms codebases into interactive, queryable maps. HCE uses tree-sitter to parse code into a hypergraph — where a single edge groups multiple related entities (caller, callee, argument types, return type) as one relationship, rather than decomposing them into pairwise connections that lose information.

Supports 9 languages. Provides MCP tools so AI agents like Claude can explore codebase structure in milliseconds without consuming LLM tokens for parsing. Captures seven relationship types: CALLS, IMPORTS, DEFINES, INHERITS, SIGNATURE, RAISES, and DECORATES.

The research behind HCE showed that combining structural traversal with text search finds things neither approach finds alone — with the effect scaling dramatically on larger codebases.

Python Hypergraphs AST Analysis MCP Tools tree-sitter

View on GitHub → Read the Article →

Ariadne Thread

An open source document extraction and retrieval pipeline that saves AI agents 100–200x in token costs. Converts 20+ document formats into clean Markdown, chunks intelligently, embeds for semantic search, and stores everything in Postgres with pgvector.

Deploy once on Railway (free tier works), connect Claude Code with one CLI command, and your agents get six tools for document ingestion, search, and retrieval. Any MCP-compatible agent system — Open Brain, OpenClaw, Cursor — connects the same way.

Basic extraction is free via MarkItDown. The full pipeline costs pennies per document. The expensive model never touches raw documents.

Python Document Pipeline pgvector MCP Tools Semantic Search MarkItDown

View on GitHub → Read the Article →

More projects coming soon. In the meantime, you can find my other work on GitHub.