I build production agentic systems — multi-agent pipelines, voice agents, retrieval over large corpora, and the evals that keep them working once they're live. Currently building for startup clients at Voxel Labs.
Before this I spent 3 years at Stripe on the startup team, leading the Y Combinator partnership and the Banking-as-a-Service startups initiative. Before that I founded an AI bookkeeping company (YC S16).
In my spare time I'm an avid hackathon enthusiast. Over the past few years I've participated in over 50, won ~30, and have hosted a few myself.
A few of my open-source projects
| Project | What it does | |
|---|---|---|
| claude-cost-teardown | One agent taken apart call by call — what a run actually costs, which fix is worth doing, and in what order. Every figure in the README regenerates from the results CSV; none are typed by hand. | source · writeup · talk |
| deskwork | Grounded Claude computer use that uses on-the-fly retrieval. | source · writeup |
| codefleet | A file lock for parallel Claude coding agents. The second one to reach for a file is refused at the tool call, before the write lands. | source · writeup · demo |
| datagraph | Royalty splits for AI answers — Shapley attribution over the providers whose records actually shaped the response. | source · writeup · demo |
| tollgate | Auth-and-capture for AI API calls, settled on-chain. | source · writeup |
| sketch-code | Generates HTML from a hand-drawn wireframe. Written in 2018 on TensorFlow 1.x, when solving this meant training a model for the task. Still cloned every week. | source · writeup |
| llm-full-stack-tutorial | A working retrieval application, and the two-part series that builds it from scratch. | source · writeup |