A2A-MediFlow AI: Intelligent Prescription & Risk Orchestration:
Inspiration
Healthcare workflows often involve fragmented systems—clinical notes, lab reports, prescriptions, and compliance policies exist in silos.
This project was inspired by the need to: unify clinical decision support automate prescription generation integrate risk detection with real-time patient data ensure compliance with healthcare policies like HIPAA
What it does
Generates structured prescriptions based on clinical conditions Calculates risk factors (e.g., hypertension, cholesterol risks) Uses FHIR APIs to retrieve patient context securely Applies policy-aware logic for safer outputs Demonstrates Agent-to-Agent orchestration for modular AI systems
How we built it
Architecture A2A (Agent-to-Agent) flow: GeneralChatAgent → External Agent → Tools → LLM Components PrescriptionServiceAgent generate_prescription_tool calculate_risk_factors_tool LLM gemini-2.5-flash for reasoning + formatting FHIR Integration Secure API-based patient data retrieval Middleware Handles authentication (API key) Injects FHIR context App Layer app.py, app_factory.py ngrok for external access
Challenges we ran into
FHIR data inconsistency: Multiple observations with conflicting values (e.g., triglycerides) Sorting issues: FHIR _sort not reliable across servers Agent orchestration complexity: Managing tool outputs + LLM reasoning cleanly Balancing determinism vs LLM: Avoiding hallucination in prescriptions Security concerns: Handling PHI safely with API-based context
Accomplishments that we're proud of
Built a modular A2A healthcare agent system Integrated FHIR with AI agents Achieved deterministic + LLM hybrid design Added risk-aware prescription generation Created a scalable multi-agent architecture pattern
What we learned
- FHIR data requires client-side validation + sorting
- LLMs should format, not decide critical medical data
- Tools provide ground truth, LLM provides presentation
- Agent orchestration improves scalability + separation of concerns
- Healthcare AI must prioritize: accuracy compliance explainability ## What's next for A2A MediFlow AI Add Drug Interaction Checker Agent Expand to multi-patient workflows Add clinical guideline validation Implement FHIR write-back (MedicationRequest resource) Deploy on OpenShift / Kubernetes with AI Orchestrator ------------------------------------------------------------------------------------------------------------------
Project 2-MCPHealthContextOctopus AI: MCP-Powered Patient Insight Engine:
Inspiration:
Healthcare data is often scattered across multiple systems—lab results, vitals, and clinical records are not unified. This project was inspired by the need to: create a single source of truth for patient health context enable dynamic tool discovery via MCP integrate FHIR-compliant healthcare data with AI agents reduce manual effort in interpreting patient data
What it does
Retrieves patient data using FHIR APIs Exposes clinical tools via MCP server Aggregates: Blood Pressure Cholesterol Triglycerides Builds a structured patient health context Enables A2A + MCP hybrid architecture Uses LLM (Grok / Gemini) for reasoning and summarization
How we built it
Core Architecture Custom Agent → MCP Server → Tools → FHIR API pipeline
Pipeline Components Agent Layer PatientHealthContextAgent System prompt-driven context builder MCP Server (mcp-health-context-server) Tool mapping layer Exposes: GetPatientBloodPressure GetPatientCholesterol GetPatientTriglycerides FHIR Layer fhir_client.py API-based retrieval using secure token Utilities mcp_utilities.py → response formatting mcp_instance.py → MCP server initialization API Layer FastAPI ngrok for external exposure
Context Handling
FHIR context injected via middleware / environment
Challenges we ran into
- Unable to fetch latest data uploaded
FHIR returned older observations instead of latest
_sort was unreliable across server implementations
Required manual sorting using effectiveDateTime
- MCP Tool Mapping Complexity Mapping tools dynamically to MCP server required careful design Ensuring consistent input/output contracts
- Data Consistency Issues Multiple records for same metric (e.g., triglycerides) Needed validation and filtering (final vs preliminary)
- Context Propagation Passing FHIR context correctly across agent → MCP → tools
Accomplishments that we're proud of
Built a custom MCP server for healthcare tools Successfully integrated FHIR APIs with AI agents Enabled dynamic tool usage via MCP skills layer Created a scalable multi-agent + MCP architecture Solved real-world issue of clinical data inconsistency
Log in or sign up for Devpost to join the conversation.