A statistical question travels to each site; the records stay there. Built on Flower, CohortLens combines local results and uses an LLM investigator to propose follow-up analyses when the sites disagree. Every follow-up needs human approval.
The demo uses public cancer cell-line data split across three emulated sites. Analysis contracts, query budgets, and a hash-chained ledger make each round inspectable.
Python / Flower / NumPy / LLM tools
Public-source research · Evidence-backed extraction
Lead Pipeline
A pipeline that turns public sources into a ranked list of Berlin property managers. It discovers companies, collects websites and documents, and uses the Claude Agent SDK to extract facts with supporting source quotes.
Validation checks the extracted evidence before scoring and export. SQLite stores progress so interrupted runs can resume; a public dashboard presents redacted results.