Jerry Zhang

Jerry Zhang is cofounder and CEO of Lemma, a San Francisco startup that monitors AI agents in production, surfaces silent semantic failures, and helps teams fix them. Lemma scans traces for cases where an agent looks successful but did the wrong thing, points at root causes in the agent chain, and can open a pull request with an improved prompt. Zhang and cofounder Cole Gawin are in the YC Fall 2025 batch.

Leaving USC classes to build together

Zhang and Gawin met as freshmen at USC and kept building instead of sitting through every class. Before Lemma they engineered at AI-native startups Tandem (healthcare) and Chipstack (chip-design agents). At both companies they saw eval setups that looked like clunky Retool dashboards and manual prompt tweaking, then built internal systems that automated evaluation and the error-driven feedback loop.

Continuous learning instead of one-off evals

Lemma's Launch YC writeup frames the problem as production drift: prompts and agents degrade as real users invent new edge cases, and engineers lose days digging through logs. Their product closes that loop - detect failed outcomes from live traffic, run targeted prompt optimizations, and either open a PR or serve the improved prompt from Lemma's dashboard - alongside tracing, drift detection, and regression alerts.