Who Sells Data 216 vendors · updated Nov 18, 2026Talk to the founder

Raycaster

Raycaster offers evaluation programs, evaluation and training data, and reproducible environments for testing and training AI agents on professional work in documents, spreadsheets and other tools, and says it builds them with domain practitioners. Its own published benchmark is Biopharma Bench, which tests agents on pharmaceutical regulatory and manufacturing (CMC) work. It also offers its own document and spreadsheet agents.

Suggest an edit
Headquarters
New York, New York, United States
Founded
2025, per its YC and LinkedIn pages (YC lists it in the Fall 2024 batch)
Headcount
1-10
Total funding
Not confirmed
Latest round
Y Combinator, Fall 2024 batch (amount not confirmed; other investors and any later funding not covered)

Trust & data

How the vendor protects buyer material, where its data comes from, and where it can run.

Trust center

Sources:raycaster.ai/securityraycaster.ai/privacy-policy

Quote
Enterprise security for regulated document work.
Data sources
Expert or contractor-createdSynthetic / generatedPublic / open dataBuilt from customer data

Sources:raycaster.ai/datahuggingface.coeval.raycaster.airaycaster.ai/domains

Quotes
  • Expert or contractor-created: Experts provide representative work, source material, and edge cases.
  • Synthetic / generated: Havenor was built from the ground up as a synthetic showcase company.
  • Public / open data: we publish two open companion showcases in full: Havenor Therapeutics (synthetic) and the Broad Institute (public FDA record)
  • Built from customer data: Bring the tasks, tools, files, and decisions your agents need to navigate
Stated confidentiality practices
Raycaster says the environments behind its main benchmark scores are grounded in real sponsor filings and private records from benchmark partners, that those records stay sealed, and that partner identities, raw packets and holdouts stay qualified or under NDA unless the contributor approves release. Its general policies, which don’t separately address data or environment programs, say it doesn’t use Customer Content to train generalized foundation models and encrypts it in transit and at rest, though it may use anonymized, de-identified or aggregated information from service use for any lawful purpose unless the customer opts out, and opting out only applies going forward.

Sources:raycaster.ai/expertseval.raycaster.aihuggingface.coraycaster.ai/privacy-policyraycaster.ai/terms-and-conditionsraycaster.ai/security

Quote
Partner identities, raw packets, and holdouts stay qualified or NDA unless you explicitly approve.
Not publicly listed
SOC 2, Other certifications, Deployment, Exclusivity

What they sell

Named customers
Not confirmed

Every change to these facts is kept in the public claims log. Record co_0687440ea053.