I'm a standup comedian in New York, and I do open-source research in machine learning: program synthesis, LLM-guided coding and agentic software engineering.
I write and perform comedy. I like food, I like jokes, I like learning, and that's my comedy.
I created and produce Baklava Comedy, with fresh baklava at the door. I built Library of Comedy, an archive of standup specials and a live show where we give away books.
I worked on Description2Code for OpenAI's Requests for Research in 2016 and built Hopper, a software engineering assistant, in 2017.
Today I work on program synthesis and coding agents. With the Sutro Group, I ran coding agents that reimplemented 53 of Hinton's experiments and 58 of Schmidhuber's papers in pure NumPy, with each implementation documenting what reproduced and what didn't.
Twenty years building software. Two companies founded, CTO twice, Head of ML twice. My work has included reinforcement learning for HVAC at Roby, founder matchmaking at On Deck, and NEAR Atlas at the NEAR Foundation. I co-founded KayOS while I was building My Pixel Pilot, self-hosted project management run by LLMs.
These days I'm building the infrastructure around coding agents: Chimera, staxtrace, and Novalis.
Penetration testing, application security and LLM red teaming. In our gpt-oss red-teaming work, the first problem was the measurement: the baseline detector was counting proper refusals as successful attacks. We built a better detector before trusting the results.
I've taught machine learning and linear algebra to Georgia Tech OMSCS students through Codementor, and led a deep learning study group in San Francisco. At AgenticEngineer.ING, I teach working with coding agents.
The technical work is at yad.codes and programsynthesis.pub. The blog here is for the other interests: comedy, maps, languages, self-learning, and internet rabbit holes.