Praxis

Inspiration

Medication harm often appears quietly: one concerning event at a time, spread across many charts. We wanted to build something that helps safety teams see the pattern earlier.

What it does

Praxis scans an entire FHIR workspace for emerging drug-event clusters. It identifies affected patient groups, explains why a signal is credible, and produces FHIR-native review artifacts like DetectedIssue, Composition, Provenance, and AuditEvent resources.

How we built it

We built a Prompt Opinion MCP server around a deterministic detection pipeline: Poisson z-score, Bayesian online change-point detection, CUSUM, and Benjamini-Hochberg false-discovery control. LLMs explain the results; statistics decide what fires.

Challenges we ran into

The hardest part was making the system workspace-scale instead of patient-scale, while keeping outputs auditable and clinically grounded. We also had to tune the demo data and evaluation harness so the signals were meaningful, not just impressive-looking.

Accomplishments that we're proud of

Praxis turns raw FHIR data into a clinical safety review queue. It works through natural language, produces signed FHIR outputs, and keeps the signal logic reproducible.

What we learned

Healthcare agents need more than good answers. They need provenance, reviewability, and clear boundaries between statistical evidence and language-model explanation.

What's next for Praxis

Next, we want to connect Praxis to real-world safety workflows: longitudinal monitoring, richer baseline calibration, and direct integration into pharmacist and quality-team review queues.

Built With

• promptopinion