Zero
Eval

Jonathan Chavez and Sebastian Crossa met during their first year of college in Mexico. Seven years later, both were deep in the machinery of AI infrastructure - Jonathan on the LLM observability team at Datadog, watching enterprises struggle to understand why their models kept failing; Sebastian as a founding engineer building email at Micro (backed by a16z) and before that at Atrato (YC W21).

They had seen the same problem from different angles: companies build an AI agent, ship it, and then have no reliable way to know why it is performing badly or how to fix it. The evaluation tooling either doesn't exist, requires a small army of data labelers, or produces static judges that degrade the moment your production data drifts from your test data.

Before applying to YC, they built a side project together - llm-stats.com, an LLM leaderboard that quickly grew to 60,000 monthly active users and a third of a million unique visitors. It was proof they could build things people actually wanted. ZeroEval is the serious version of that instinct applied to a much harder problem.

That single sentence explains everything about the company's positioning. ZeroEval isn't trying to make the world's best AI model. It's building the feedback infrastructure that makes your model - whatever it is - incrementally smarter every day without you having to manually audit thousands of outputs.

AI agent evaluation is a genuinely hard problem. A chatbot that answers questions is one thing. An agent that makes dozens of tool calls across a long, multi-turn session to accomplish a complex task is another. There is no simple accuracy metric. The failure modes are subtle, numerous, and deeply task-specific.

ZeroEval attacks this in three stages: observe, evaluate, fix - and then do it automatically.

Most AI evaluation startups are built by researchers who understand benchmarks but have never shipped a production system under pressure. ZeroEval's founders are operators first.

Their YC partner is Jon Xu. The batch: Summer 2025. The location: New York City, which has quietly become one of the better places to build B2B AI infrastructure - closer to the enterprise customers writing the big checks.

The AI agent wave of 2025 created a new category of operational headache. Companies are shipping agents that do things - book flights, write code, triage support tickets, analyze medical data - and discovering that "it worked in testing" and "it works in production" are very different statements.

The evaluation tooling industry is responding. But most solutions look like observability products with an evaluation tab tacked on. ZeroEval's bet is different: don't just tell developers what broke, automatically fix it. Close the loop from failure detection to prompt deployment without a human in the middle.

The vision is explicitly ambitious. In the founders' words, they see a future where "developers define the evaluation criteria as a starting point and errors back-propagate to find the optimal implementation." That's less an observability product than a continuous training loop for production software.

The MCP server integration is an underrated detail here. When ZeroEval publishes a server that lets coding agents like Claude Code query their own evaluation data to understand where they fail - AI agents using an AI evaluation tool to get smarter - that's a glimpse of the feedback loop the company is trying to institutionalize.

DoorDash, Datadog, Hugging Face, and Harvard Medical School all trusted ZeroEval early. That's a remarkably diverse set of use cases for a two-person company from a single YC batch. Delivery logistics, developer tools, ML infrastructure, and medical AI - the common thread is agents making consequential decisions where getting it wrong has a cost.