Med-Triage Voice — A2A Agent + FHIR Intake

• 

  Med-Triage Voice web interface for capturing patient audio

• 

  Transcribing patient audio in real-time

• 

  Displaying the processed text transcription

• 

  Extracting symptoms and clinical reasoning via Gemini AI

• 

  Structuring the data into a compliant FHIR R4 JSON format

• 

  Generating ESI severity levels and actionable medical advice

• 

  Testing the Med-Triage Assistant on Prompt Opinion Launchpad

• 

  Seamlessly generating FHIR triage assessments via A2A chat

• 

  Med-Triage Assistant published live on Prompt Opinion Marketplace

Inspiration Emergency room intake is a bottleneck that costs lives. Nurses spend 10+ minutes manually typing patient symptoms before a doctor even sees them.

What it does Med-Triage Voice listens to a patient describe their symptoms, transcribes the audio using Whisper, extracts a structured clinical record via Gemini AI, and broadcasts its capabilities as an A2A Agent on the Prompt Opinion network — in under 10 seconds.

How we built it FastAPI backend + faster-whisper for transcription + Google Gemini for clinical extraction + a custom A2A Agent configured with a SKILL.md package + FHIR R4 Bundle assembler with SNOMED-CT and LOINC coded resources. Deployed seamlessly on Render via Docker.

Challenges Getting Whisper to run on cloud servers without triggering HuggingFace rate limits. Solved by pre-downloading the Whisper model directly during the Docker build phase. We also had to navigate the strict .zip folder structure and YAML specifications required to publish our agent on the Prompt Opinion marketplace.

What we learned FHIR R4 bundle constraints (bdl-1), Prompt Opinion's A2A Agent architecture, and prompt engineering for reliable structured JSON from LLMs.

Built With

• a2a-protocol
• docker
• fastapi
• faster-whisper
• fhir-r4
• google-gemini
• prompt-opinion
• python
• render
• snomed-ct