Luveo Health Compliance Copilot
Extending and evaluating an AI copilot that translates structured healthcare-compliance results into clear, actionable explanations.
Luveo Health · June 2026 – Present · San Diego, CA
Interactive walkthrough
Synthetic demo. Trace a compliance result from event to explanation. No patient data. Prefer full-screen if the embed feels tight.
The problem
THE PROBLEM
Compliance needs explainable structure
Healthcare organizations must check clinical actions against interoperability standards and compliance rules. Luveo builds an AI copilot that reasons over these scenarios; rule execution, explanations, and the clinical sandbox needed extension and validation so outputs were correct and explainable.
MY ROLE & ACTION
What I worked on
- · Mapped and extended architecture normalizing Epic, Pyxis, and Omnicell-style events against pharmacy rules.
- · Built end-to-end demo integration connecting React, FastAPI, HL7 v2.5, and FHIR R4 across eight medication scenarios.
- · Instrumented the LLM explanation layer with Langfuse tracing.
- · Strengthened grounding with structured workflow and regulatory schemas.
THE RESULT
Current outcome
- · Eight synthetic scenarios return structured Pass, Warning, or Fail evaluations.
- · End-to-end AI behavior inspectable through Langfuse traces.
Architecture
- 01
Synthetic event
HL7 / FHIR / dispensing payload
- 02
Rules engine
Deterministic Pass / Warning / Fail
- 03
FastAPI case view
Structured result + evidence
- 04
Explainer agent
Grounded natural-language guidance
- 05
Langfuse
Trace prompts, tools, failures
What I personally contributed
- Mapped and extended architecture normalizing Epic, Pyxis, and Omnicell-style events against pharmacy rules.
- Built end-to-end demo integration connecting React, FastAPI, HL7 v2.5, and FHIR R4 across eight medication scenarios.
- Instrumented the LLM explanation layer with Langfuse tracing.
- Strengthened grounding with structured workflow and regulatory schemas.
Technical concepts
Click a card to expand.
Why it mattered here
Near compliance decisions, free-form generation alone is hard to trust. The score must be grounded.
How I used it
Rule engine emits structured evidence; an explainer agent turns it into actionable guidance for clinicians/ops.
Technical detail
FastAPI case view exposes typed results; React sandbox consumes authenticated endpoints.
What can go wrong
Phase 1 uses synthetic payloads — architecture validation, not production regulatory guidance.
Limitations
Phase 1 uses synthetic data for architecture validation. Not a deployed clinical product and not validated regulatory guidance.
What I learned
Applied AI work is systems reasoning — following a result across payloads, rules, APIs, and UI — and evaluating whether a system can explain itself.