Snorkel AI

Snorkel AI describes itself as a frontier AI data lab. It builds specialized training data, benchmarks and evaluation environments that help frontier models and AI agents perform in difficult, high stakes domains. The company was founded out of the Stanford AI Lab and has a research team that publishes its methods. Its approach has three repeating steps. Evaluate measures model behavior against task specific benchmarks in realistic environments with programmatic pass or fail checks. Curate runs rubric guided pipelines with calibrated domain experts, including environments with the tools, documents and verifiable reward signals agents are tested against. Refine analyzes disagreements and failures and updates rubrics and benchmarks. The company says it covers more than 1,000 expert level domains and measures reviewer calibration instead of assuming it. The customers are frontier labs and AI teams. Snorkel also runs public leaderboards and benchmarks, including computer use, workplace agents, software engineering, continual learning and science tasks, plus a grants program for open benchmarks. Its platform, research and embedded delivery model combine to produce datasets, environments and evaluation frameworks. Pricing is not published. The site directs visitors to request dataset samples or talk to the team.