Clinical Intelligence Assistant
Can an AI assistant answer questions from synthetic patient records while showing its evidence—and admit when something isn't documented?
FHIR Patient/001 · Sample only
“Summarize this patient's chronic conditions.”
Sources: Condition/12 · Condition/18
Inside the build
The problem
Patient data is spread across structured clinical resources. A useful assistant must retrieve relevant evidence, produce accurate summaries, and make its sources easy to inspect.
The goal
Build a synthetic-data prototype that answers chart questions using explicit evidence, distinguishes unknown facts from known facts, and supports reproducible evaluation.
Architecture, v0.1
Streamlit UI → FastAPI → validated retrieval tools → synthetic FHIR resources and indexed clinical context → hosted LLM → source verification → human review.
├── app/ # FastAPI and tools
├── ui/ # Streamlit interface
├── data/ # Synthetic FHIR fixtures
├── evals/ # Grounding and regression cases
└── tests/ # Python unit tests
What gets tested?
Does every clinical claim have source evidence?
Does the system say when a record doesn't support an answer?
Does it retrieve the right patient and resource?
Target: 30 reproducible synthetic cases. Metrics and pass rates will be reported after implementation—not invented in advance.
The build journal
Generate and render synthetic longitudinal records.
Connect questions to explicit retrieved records.
Test quality and publish a demonstration.
Try the concept.
This is a deterministic frontend mock demonstration. It doesn't call a language model, access real patient records, or represent a clinically validated system.
Patient 001
Choose a sample question: