CleveCare
Seven-Agent Healthcare AI
A multi-agent AI triage system built for the Cleveland Clinic Abu Dhabi Hackathon. Seven specialized agents coordinate to handle patient intake through to a clinical summary, reducing administrative burden on nursing staff.
The Problem
Nursing staff at large hospitals spend a significant share of their shift on intake, triage, and documentation instead of direct patient care. The "Code to Cure" challenge asked teams to improve patient access to the correct healthcare provider.
The Challenge
Triage is high-stakes and multi-step: it requires structured intake, symptom classification, urgency scoring, medical-history context, and a clean clinical summary — all while remaining explainable and safe. A single monolithic prompt could not reliably handle this pipeline.
Architecture
- Orchestrator built on LangGraph coordinating seven specialized agents as a directed graph.
- Intake Agent — captures and normalizes patient-reported information.
- Symptom Classification Agent — maps free-text symptoms to structured categories.
- Urgency Prioritization Agent — assigns a triage severity score.
- Condition Cross-Reference Agent — RAG over a medical knowledge base.
- Medical History Integration Agent — merges prior context into the assessment.
- Risk Flag Detection Agent — surfaces red-flag conditions for escalation.
- Clinical Summary Agent — produces a structured, sign-ready summary for staff.
Workflow
- 1Patient input is captured and normalized by the intake agent.
- 2Symptoms are classified and cross-referenced against a retrieval knowledge base.
- 3Urgency and risk flags are computed in parallel and reconciled by the orchestrator.
- 4A final clinical summary is generated and handed to nursing staff for review.
Engineering Decisions
- →Chose a graph of specialized agents over one large prompt for reliability and explainability.
- →Used RAG to ground condition cross-referencing in a controlled knowledge base rather than model memory.
- →Kept a human-in-the-loop review step — the system assists, it does not diagnose autonomously.
- →FastAPI backend for a clean, typed contract between the agent graph and the demo UI.
Results
Lessons Learned
Decomposing a hard clinical workflow into narrow, testable agents made the system dramatically more reliable and easier to reason about than a single model call — and made the demo credible to healthcare professionals.
Future Improvements
- Formal evaluation harness with clinician-labeled cases.
- Tighter EHR integration and structured export.
- Bias and safety auditing across demographic groups.