A deeper look at projects in AI agents, local language models, browser automation, and ML research. Each one is a puzzle I enjoyed solving.
Featured
Claude Code in 100 Lines
Shipped2026
A from-scratch reimplementation of how agent frameworks like Claude Code work — written in ~110 lines of Python. Every agent loop, tool dispatch, subagent, skill, and memory pattern is exposed and readable, running locally via Ollama with no Anthropic API calls.
The full agent loop in a single readable file (llm.py)
Tool schemas auto-derived from Python signatures — no JSON to maintain
Progressive-disclosure skills and memory: tokens paid only when needed
Subagent spawning with bounded recursion to isolate noisy sub-tasks
BrowserControl
Active2024
A vision-first browser automation MCP server for AI agents. Instead of brittle CSS selectors, agents receive annotated screenshots with numbered elements and simply say "click 5." This makes browser automation far more robust for dynamic, real-world interfaces.
Annotated screenshots with numbered UI elements
MCP server compatible with Claude Desktop and other clients
Works on dynamic SPAs without brittle selectors
Published on PyPI with open-source community usage
More projects
Cutiepie
Shipped
A VS Code color theme built on exactly two accents over true black. Powder pink carries structure (functions, classes, tags, JSON keys), pale sky blue carries content (strings), and everything else stays greyscale. Ships as a dark and a light variant.
Two-accent palette: pink for structure, sky blue for content
Matched dark and light variants from one set of color decisions
Diff and git colors rebuilt around the palette, with errors kept deliberately rosy
Previews and palette swatches generated from the theme JSON, so the docs cannot drift
Want to see more?
Most of my work lives on GitHub, including experiments that never became full projects.