Lumen

Inspiration

Clinical AI is moving fast, but most systems still produce text without proving whether the recommendation is safe for the patient in front of them. We wanted Lumen to be the verification layer: something that checks before anyone acts.

What it does

Lumen verifies patient-specific clinical recommendations using the selected FHIR chart. It returns PROVEN_SAFE, PROVEN_UNSAFE, or UNDETERMINED, with cited guideline logic, simulation context, safer alternatives, and FHIR-native audit artifacts like Composition, Provenance, and AuditEvent.

How we built it

We built a Prompt Opinion MCP agent around a verification pipeline: FHIR state extraction, encoded guideline theorems, Z3 formal checking, digital-twin trajectory simulation, value-of-information ranking, adversarial alternatives, and signed FHIR provenance. LLMs can propose or explain; they do not decide the safety verdict.

Challenges we ran into

The hardest part was making the system strict without making it brittle. Lumen has to abstain when data is missing, cite the exact rule when unsafe, and still fit naturally into a clinician’s Prompt Opinion workflow.

Accomplishments that we're proud of

Lumen turns a free-text recommendation into a reviewable safety certificate. It works through natural language, grounds itself in FHIR, produces auditable outputs, and keeps clinical safety decisions outside the LLM.

What we learned

Healthcare agents need more than confident answers. They need proof, provenance, uncertainty handling, and clear separation between language generation and clinical decision logic.

What's next for Lumen

Next, we want to expand guideline coverage, improve VOI calibration, support richer write-back workflows, and integrate Lumen as a verification layer for any clinical AI agent.

Built With

• promptopinion