Snorkel AI, Inc. is an American company specializing in data development for artificial intelligence. Based in California, it markets Snorkel Flow, a platform designed for teams building machine learning applications. Its guiding principle: reducing reliance on manual annotation, which is often time-consuming and costly, by turning domain specialists’ knowledge into reusable, controllable labeling mechanisms.
From Stanford research to a company
Founded in 2019, Snorkel AI builds on a research project launched at Stanford around weak supervision. This method uses imperfect sources of information — rules, dictionaries, knowledge bases or predictions from other models — to produce training data. Rather than requiring a human label for every example, it combines these signals and accounts for their disagreements.
The open-source Snorkel project helped spread this approach within the machine learning community. The company, whose co-founders include Alex Ratner, subsequently developed a commercial offering to apply it to organizational constraints: collaboration between experts and developers, data management and integration into model development workflows.
Scaling up data work
Snorkel Flow allows users to define labeling functions, meaning rules or programs that automatically assign categories to examples. A team can, for instance, detect certain terms in documents to identify their subject or extract useful information. The platform then helps combine these signals, train models and analyze their errors.
The benefit is not limited to speeding up initial annotation. When a business need changes or a new category emerges, teams can modify their rules and regenerate labels rather than rework the entire corpus manually. This approach is particularly relevant to text classification and the use of documents in sectors with specialized vocabulary and requirements.
With the rise of generative AI, Snorkel AI also applies its approach to preparing data suited to large language models and to evaluating those models. Its business model is based on selling software and services to businesses. The snorkel.ai website presents these offerings, along with technical resources devoted to data-centric AI.
What next?
The widespread adoption of pretrained models is shifting some of the competition toward each organization’s own data. For Snorkel AI, the challenge is to make this adaptation more reproducible while demonstrating its value compared with annotation assisted by the models themselves. Weak supervision does not, however, eliminate the need for human oversight: poorly designed rules can propagate errors or biases. The ability to measure data quality and performance on concrete business use cases will therefore remain crucial to its development.