Datalab.to

Datalab.to is a productivity tool. Datalab trains document intelligence models and offers them through an API with Convert, Extract, Segment and Eval processors. It runs as managed cloud, in a customer VPC or air-gapped, aimed at AI labs and regulated industries.

Datalab is a research lab that trains document intelligence models and sells them as an API and as deployable software. The core job is turning unstructured files such as PDFs, spreadsheets and slides into structured output that downstream systems can trust. Its customer list on the homepage includes frontier AI labs, museums and enterprises in fields where accuracy matters more than speed or price. The product is organized into four processors. Convert turns documents into markdown, HTML or JSON using the Chandra model, with word-level bounding boxes and redlines included. Extract pulls key fields using predefined schemas. Segment finds the boundaries between documents that were scanned or merged into a single file. Eval scores output quality against Datalab's own criteria or a team's custom rubrics. Processors can be composed in the playground, then promoted as a versioned pipeline to production, with continuous evals meant to flag regressions as models are updated. Deployment is a notable differentiator. Teams can use the managed cloud (sign up, get an API key, pay as you go, with a free tier), run Datalab inside their own VPC on AWS, GCP or Azure, or install it fully air-gapped on-premises with full model weights. Because Datalab owns its models, the network choice is left to the customer. The company lists SOC 2 Type II, custom BAA and DPA terms, and dedicated support on the higher tiers. The homepage also points to open-source work: Marker and Surya for lightweight OCR in 90+ languages, and Chandra as the flagship model. Published benchmark charts compare the Datalab API against Chandra variants and Gemini 2.5 Flash. Datalab suits AI teams building training corpora, and regulated industries such as finance, healthcare and insurance that need audit-ready extraction. Compared with general-purpose multimodal LLMs or cloud OCR services, it positions itself as a specialist with a stronger focus on precision and deployment control. Pricing beyond the free tier is not detailed on the homepage.