Live MCP healthcare orchestration server running on Railway with Gemini-powered clinical workflows
FHIR-integrated clinical intelligence tools exposed through MCP architecture

Inspiration

Hospital discharge is one of the most failure-prone moments in healthcare.

Nearly 1 in 5 patients are readmitted within 30 days after discharge, often because symptoms worsen unnoticed, medications are misunderstood, follow-ups are missed, or clinical information becomes fragmented across systems.

CareFlow AI was built to address this gap using interoperable AI orchestration instead of isolated healthcare chatbots.

The goal was to create a SHARP-compliant MCP clinical intelligence platform capable of reasoning over structured FHIR patient context in real time.

What it does

CareFlow AI is an MCP-native healthcare orchestration platform that exposes AI-powered clinical workflows through interoperable tools.

The system combines:

Gemini-powered clinical reasoning
FHIR R4 interoperability
MCP tool orchestration
SHARP-aligned healthcare context propagation
Structured JSON clinical outputs

CareFlow AI currently exposes 5 interoperable clinical intelligence tools:

1. analyze_symptoms

AI-powered symptom triage using symptom combinations, patient context, and vital signs to identify dangerous clinical deterioration patterns.

Unlike static rules engines, the system reasons across symptom relationships and contextual patient burden.

2. detect_care_gaps

Analyzes FHIR patient history to identify overdue screenings, monitoring gaps, follow-up failures, and unresolved care coordination issues.

3. compute_readmission_risk

Predicts 30-day readmission risk using:

active conditions
medications
prior admissions
transportation access
social determinants of health
home support availability

The system generates prioritized intervention recommendations instead of only numeric scores.

4. generate_handoff_note

Automatically generates structured SBAR clinical handoff documentation from patient context and workflow state.

This reduces manual documentation overhead during discharge transitions.

5. medication_reconciliation

Detects medication conflicts, duplication risks, adherence concerns, and transition-of-care medication issues during discharge planning.

How I built it

Backend Infrastructure

Python
FastAPI
FastMCP
Uvicorn ASGI server

AI Layer

Google Gemini API
Structured clinical JSON generation
Prompt-based reasoning orchestration

Healthcare Interoperability

FHIR R4
HAPI FHIR sandbox
SHARP-on-MCP context propagation
MCP Streamable HTTP transport

Deployment

Railway cloud deployment
Production ASGI runtime
JSON-RPC based MCP communication

The platform was architected as a modular MCP server where every clinical workflow operates as an independent interoperable tool.

Challenges I ran into

One of the hardest parts was implementing MCP streamable transport correctly while maintaining SHARP-compatible healthcare context propagation.

Additional challenges included:

handling FHIR interoperability cleanly
generating consistent structured clinical JSON
integrating Gemini reasoning reliably
debugging Railway deployment/runtime issues
designing healthcare-safe tool execution patterns
exposing MCP capabilities correctly during initialization

Accomplishments that I'm proud of

Built and deployed a fully working MCP healthcare server
Implemented 5 interoperable clinical intelligence tools
Added SHARP-aligned FHIR context propagation
Successfully deployed production infrastructure on Railway
Integrated Gemini-powered clinical reasoning
Added medication reconciliation and social determinant-aware readmission analysis
Implemented structured AI orchestration instead of chatbot-only interaction

What I learned

This project demonstrated how future healthcare AI systems will likely evolve beyond standalone chatbots into interoperable clinical orchestration platforms.

I learned: