I work on agentic AI at Abridge, building ambient AI for clinical decisions. Before that, agents, evaluation and LLM inference at LinkedIn — mostly the unglamorous parts of getting agents to survive contact with real users. Before that, recommender infrastructure, and a liberal arts degree in computer science, mathematics and statistics with two honors theses I still think about.

I write here about what breaks: single-step evals that pass while the twenty-step agent fails, papers that turn into exactly one engineering decision, and formal methods that turned out to be load-bearing for AI.

I also maintain a few open source packages — ccrvam, chipfiring and QuantileFlow.

My operating principle is one borrowed phrase: stay green. Curious enough to ask the obvious question in a room full of experts, and hands-on enough to go test the answer that night.