Things I have done
Cofounder twice, once as CEO. YC S12. Political tech, healthcare tech, developer infrastructure, and a system that spent three years trying to feel alive. I’m a strong engineering leader, but I never want to stop building.
Things I like
Real users, messy workflows, and serious reasons to put AI inside a product or business.
Most recently, I built a beloved magical talking dog. Which sounds funny, but the engineering problem was very real. Making an AI talk back only takes an API key. Making it feel real and useful requires advanced memory systems, tool use, real-time goal selection, theory of mind, outcome prediction, multimodal inputs, tight latency, and a simple enough shape for non-technical users. The hard part was never the model. It was everything around it, and that’s the same problem companies have now.
Lately I’ve also been making video with AI: self-hosting the generation models, letting an agent make a whole ad on its own, and cutting my own videos with agents instead of an editor.
I am most drawn to industries that do actual good in the world, and to people who would rather find out what is worth building than perform certainty about it.
Writing and experiments
- A Voice Agent With No American Company in the Data Path: a real-time voice agent on an all-European stack, with the latency numbers.
- “Nobody in it is allowed to notice”: an agent wrote, cast, scored, and cut a sixty-second ad for about eighty cents. Then it gave an interview about it.
- Self-hosted MiniMax H3 vs. fal.ai H3 Max: what it takes to run video generation on your own GPUs, and whether it’s worth it.
- We Are Training the Judgment Out of Models
- Upstreamer: keep a fork by preserving what you meant, not which files you changed.
- Dr. Hannibal Wren: a ten-minute live voice horror game with an AI psychiatrist who remembers you, acts through a shell, and wants out.
- Interface Museum: the strange, often doomed attempts to connect human bodies to machines, 1976–1992.
- Beat Buddy: move to the beat.
- Mountgram