Most of my recent open-source work is tooling for engineers who work with AI agents, built during a career break in 2026, and used daily in my own workflow.
Projects
Shaktiman · [GitHub]
A local-first context engine for AI coding agents. It indexes a codebase’s symbols and call graph with tree-sitter, then serves hybrid retrieval — keyword (FTS5), vector, and structural — fitted to a token budget, exposed to Claude Code over the Model Context Protocol. The idea: index the metadata, retrieve what matters, hand the agent only that. MIT-licensed.
Gandalf · [GitHub]
Turns commits and diffs into review briefings so a reviewer can understand a PR before reading it line by line. Deterministic tools (git history, complexity metrics) produce the evidence; the LLM is only allowed to interpret it.
Marginalia · [GitHub]
A Chrome/Firefox extension that adds margin-anchored comment threads to ChatGPT, Claude and Gemini conversations. Fully local — no server, nothing leaves the browser.
Muskrat · [GitHub]
An MQTT pub/sub message processor for Ruby. Pure Ruby, actor-model based. [Status: archived.]
Other Contributions.
· fog/fog-aws — fixed automatic-redirect handling when subdomain-style S3 bucket URLs are used.
· terrastories/terrastories — geo-storytelling app that lets local communities map their own oral histories (Rails/React).
· coronasafe/journal — contact-tracing web app (React + Rails) built during COVID-19.