MedMitra — AI Medical Case Management
Multi-agent system using LangGraph to automate medical case analysis — processing 500+ patient notes, lab reports (LlamaParse), and radiology images (Llama 4 vision) with 94% accuracy. SOAP notes and diagnostic suggestions in under 3 seconds.
Problem
Medical case review involves messy multimodal context: notes, labs, reports, images, and timelines. MedMitra explores how agentic systems can structure that evidence before producing a useful clinical-style summary.
Approach
- –Used LangGraph to split the workflow into specialized analysis nodes instead of asking one model to reason over every artifact at once.
- –Parsed patient notes and lab reports through LlamaParse, then paired that text evidence with radiology image interpretation.
- –Generated SOAP notes and diagnostic suggestions after evidence collection, keeping the output format predictable and reviewable.
System Diagram
Case intake
Receives patient notes, lab reports, and radiology images from the app.
Document parsing
Uses LlamaParse to normalize PDFs and clinical text into agent-readable context.
LangGraph agents
Coordinates evidence extraction, image reasoning, summary writing, and final review.
FastAPI backend
Runs the workflow, stores case state, and returns structured outputs.
Clinical output
Produces SOAP notes and diagnostic suggestions for human review.
Key Decisions
Why LangGraph
A graph makes the workflow explicit: each node owns a narrow part of the case, and the final response can be assembled from traceable intermediate findings.
Why SOAP notes
SOAP gives the model a constrained output contract, which is easier to review than a free-form medical paragraph.
Concepts Used
Outcomes
- –Processed 500+ patient notes, lab reports, and radiology images in the prototype dataset.
- –Generated SOAP notes and diagnostic suggestions in under 3 seconds.
- –Reached 94% accuracy on the measured case-analysis workflow.