Stemline RWE Intelligence lets analysts ask questions in plain English and get back governed, reproducible answers from Flatiron and Guardant data on Databricks — every step recorded, every result traceable, and a human in the loop before anything is trusted.
Six ideas explain the whole platform. The pages above go deeper on each.
We connect to Databricks with a machine identity and read data in place. No bulk data is exported.
The platform can only read the specific data it’s granted, and can only run safe read queries — never change anything.
Enterprise-licensed AI turns questions into governed queries and formats & explains results. It never owns the data or the verdict.
Every cohort and result starts as a draft. A reviewer validates it before it’s trusted or reused.
Every query, cohort, review, and chat is stored with its exact inputs — reproducible and inspectable.
Built on a mainstream, auditable tech stack. No autonomous-AI framework — we control every step.
The same real-world data, a fundamentally different way of working with it.
What happens, end to end, when an analyst asks something — every hop is governed and recorded.
e.g. “How many ESR1-positive patients had prior CDK4/6i?” — typed into the workspace.
The AI doesn’t touch data directly. It chooses one of our pre-built, validated tools (browse schema, build cohort, run a read-only query, make a chart).
The platform authenticates as its service principal, and runs a read-only query on the SQL warehouse — compute stays inside Databricks.
Aggregated results come back; the AI turns them into a clear table, chart, and plain-language summary with caveats.
Cohorts and feasibility reports are saved (de-duplicated), starting as Draft.
A reviewer marks it Validated, Needs-changes, or Rejected. Only validated definitions are reused.
The exact query, AI model + prompt version, and data version are recorded — ready for inspection or a regulator.