Inspiration
- Dense Clinical Jargon: Discharge summaries overwhelm patients with technical terminology, increasing the risk of medication errors and avoidable readmissions.
- Administrative Burden: Manually auditing every departing patient's file against mandatory hospital discharge protocols is highly tedious and error-prone.
What it does
- Extraction & Compliance Auditing: Securely structures raw clinical notes into clean JSON and cross-references the care plan against hospital rules to automatically flag missing protocols.
- Empathetic Literacy Translation: Translates complex jargon into an easy-to-understand, 8th-grade reading level daily routine while explicitly isolating emergency "Red Flag" warnings.
How we built it
- Custom MCP Backend: Built an independent, asynchronous Python MCP server (
FastMCP) featuring deterministic, hardcoded tools to completely eliminate LLM hallucinations. - State Machine Orchestration: Configured a central Orchestrator Agent with authoritative default trigger behaviors to lock in sequential tool execution (Extract $\rightarrow$ Audit $\rightarrow$ Translate) without manual user intervention.
- Optimized Containerization: Packaged the backend in a lightweight
python:3.12-slimDocker container to ensure seamless wheel installations, unbuffered logging, and stable cloud deployment.
Challenges we ran into
- FHIR API Restrictions: Overcame
403 Forbiddenblocks on restricted database endpoints by pivoting our Python tools to extract structured data directly from the active workspace text. - Agent Handoff Loops: Solved early LLM handoff loops and tool hallucinations by baking strict negative constraints ("Role Denial") and automatic triggers directly into the master prompt.
Accomplishments that we're proud of
- Frictionless Automation: Achieved an incredibly smooth, one-click user experience where simply asking to "Process this patient" autonomously drives the entire multi-tool workflow.
- Hybrid Logic Precision: Successfully paired deterministic Python guardrails (for rigorous parsing and auditing) with generative AI reasoning (for empathetic translation).
What we learned
- MCP Integration: Mastered building and securely exposing local tools via Server-Sent Events (SSE) to external cloud-agent platforms.
- Advanced Prompt Control: Learned how to enforce strict sequential state machines and override default LLM conversational hesitations using negative constraints.
What's next for Discharge Navigator
- Live EHR Webhooks: Integrating bi-directional webhooks for automated, real-time synchronization with electronic health record systems at the exact moment of discharge.
- Dynamic Rulebooks & Localization: Scaling the compliance auditor to ingest dynamic, department-specific JSON rulebooks and adding native multilingual translation for patient guides.
Built With
- fastmcp
- python
Log in or sign up for Devpost to join the conversation.