A San Francisco startup built the open-source layer that grades AI answers. Then the enterprises who ship those answers - BCG, AstraZeneca, Microsoft - started paying attention.
Every software team learned the same lesson twice. The first time was decades ago: code without tests breaks in ways you only discover in front of customers. The second time is happening right now, with AI. Companies wired large language models into support desks, medical assistants, and insurance workflows - and then realized they had no reliable way to answer a simple question. Is the thing actually any good? Confident AI, a San Francisco company from Y Combinator's Winter 2025 batch, exists to answer that question with a number instead of a shrug.
The founders are Jeffrey Ip, an engineer who did tours at Google and Microsoft, and Kritin Vongthongsri, a former Princeton machine-learning researcher. The origin story is unglamorous in the best way. Ip was building a retrieval-augmented API, hit the wall of "how do I know this is working," and found the tooling missing. So he wrote it himself and open-sourced it. That library became DeepEval, and DeepEval became the company.
The pitch developers repeat is short: DeepEval is "Pytest for LLMs." That comparison does a lot of quiet work. Pytest is the testing library Python developers already live in, so the mental model arrives pre-installed. You write assertions, you run them, you get a pass or a fail. DeepEval swaps brittle string-matching for research-backed metrics that score answers on things like relevance, faithfulness, and hallucination - turning "the new prompt feels better" into a value you can put in a dashboard and defend in a meeting.
The evaluation framework - unit-test and benchmark RAG pipelines, agents, and chatbots with deterministic, LLM-as-a-judge metrics.
Red teaming built on DeepEval. Simulates jailbreaks, prompt injection, and multi-turn exploits to surface 40+ vulnerability types.
The paid layer - dataset editor, regression catcher, tracing, alerting, scheduled attack campaigns, audit logs, and compliance reporting.
DeepEval spread the way developer tools do - bottom up, one engineer at a time. But the names that showed up are telling. Confident AI's frameworks are used inside BCG, AstraZeneca, AXA, and Microsoft, alongside a roster that stretches to Panasonic, Toshiba, Samsung, Stellantis, and Mercedes-Benz. These are consultancies, drugmakers, insurers, and carmakers - organizations for whom a confidently wrong chatbot is not a bug, it is a liability.
That customer mix is also the strategy. Regulated industries do not want a clever demo; they want evidence. Confident AI maps its red teaming output to frameworks auditors already trust - OWASP's LLM Top 10, NIST's AI Risk Management Framework, and MITRE ATLAS. It is a translation layer between "our AI seems safe" and "here is the report."
The problem Confident AI attacks is that AI development quietly skipped a step the rest of software takes for granted. Teams iterate on prompts and models by feel, ship, and hope. When something drifts - a model update, a prompt tweak, a new edge case - they often find out from an angry user. DeepEval's regression catcher is aimed squarely at that: run the tests on every change, and catch the answer that got worse before it reaches production. The company says that discipline has helped teams cut model costs by more than 70 percent and compress deployment cycles from a week or two down to hours.
Plenty of companies now sell LLM evaluation and observability - Braintrust, LangSmith, Arize, Galileo, Patronus, and others crowd the space. Confident AI's wedge is that its metrics are open source. The framework is free, transparent, and already inside your codebase; the cloud platform is what you buy when you need a team to collaborate, schedule adversarial runs, keep audit logs, and prove compliance instead of running one-off attacks from a notebook. Open source got them the users. The platform gets them the enterprise.
In March 2025, days after YC's Demo Day, Confident AI raised a $2.2M seed round and closed it in five days. Ip has written candidly about the week: an early $500K, then twenty back-to-back calls with no commitments, then a shift once an investor got interested less in the projections than in the open-source community already using the tool. By Friday they had passed their target and were asking interested investors to take smaller checks. Y Combinator led, joined by Flex Capital, Oliver Jung, Vermilion Cliffs Ventures, Liquid 2 Ventures, January Capital, and Rebel Fund.
Confident AI is not building the flashy part of AI. It is building the part that decides whether the flashy part is allowed to talk to your customers - the testing, monitoring, and governance rail underneath the applications. As more of the economy routes decisions through language models, the demand for a standard way to measure and defend those decisions grows with it. A seven-person team turned an evaluation headache into open-source infrastructure, and then into a company betting that quality assurance is the next battleground in AI. The bet is straightforward: the models will keep changing, but somebody always has to grade the homework.