I'm interested in AI products and like building things in that space.
Most of the repos here are small, self-contained projects I built to understand how LLM systems work — plus a few tools I use day to day. Everything is meant to be easy to clone and run.
From-scratch builds of the pieces behind a language model. Written to learn how things work, not for production use.
- mosaic — a byte-pair-encoding tokenizer you can train on your own text · demo
- nabla — a small reverse-mode autograd engine with a visualizer · demo
- loom — a small GPT written in NumPy, trained on Shakespeare · demo
- warren — HNSW nearest-neighbor search, the algorithm behind vector databases · demo
Practical pieces around LLM apps. Most run without any API keys.
- tinycoder — a minimal coding agent in ~570 lines of Python
- stencil — constrained decoding: guide generation with a JSON Schema · demo
- winnow — packs the most relevant context into a token budget for RAG · demo
- mend — repairs malformed JSON from LLM output · playground
- semcache — a semantic cache for LLM calls
- introspect-mcp — MCP server that serves real signatures from your installed packages
- deja — local search over your Cursor/Claude agent transcripts