MediVoice

Patients speak. Doctors arrive prepared. Voice-first medical intake powered by Amazon Nova

Comment

Inspiration

Every doctor's appointment begins the same way — "What brings you in today?" The doctor spends the first 5-10 minutes collecting basic information they could have had before walking in. Patients fill out paper forms they don't understand. Non-English speakers struggle. Elderly patients get frustrated. Time is wasted on both sides.

I asked: what if the patient could have that conversation BEFORE the appointment?

What it does

MediVoice is a voice-first medical intake assistant that replaces paper forms with a natural AI conversation before doctor appointments.

  1. Patient logs in and clicks Start Voice Check-In
  2. Amazon Nova 2 Sonic conducts a warm, empathetic intake conversation — asking about symptoms, pain level, medications, and allergies one question at a time
  3. Patient clicks End Check-In
  4. Amazon Nova Lite analyzes the full transcript and generates a structured clinical brief instantly
  5. Doctor logs into a separate secure portal and sees all today's patients with complete clinical briefs ready

How I built it

  • Frontend: React + TailwindCSS — separate portals for patients and doctors
  • Voice AI: Amazon Nova 2 Sonic via AWS Bedrock WebSocket streaming API
  • Report AI: Amazon Nova Lite via AWS Bedrock InvokeModel API
  • Backend: FastAPI (Python) with JWT authentication
  • Storage: File-based persistence for demo purposes

The most technically challenging part was implementing real-time bidirectional audio streaming with Amazon Nova 2 Sonic — capturing microphone input, converting Float32 PCM to Int16, encoding to base64, and streaming via WebSocket while simultaneously receiving and playing back the AI's audio response.

Challenges I faced

  • Real-time audio streaming: Getting the exact WebSocket event format right for Nova 2 Sonic took significant debugging — the order and structure of sessionStart, promptStart, contentStart, and audioInput events had to be perfect
  • Audio playback quality: Output audio is 24000 Hz while input is 16000 Hz — mismatching these caused garbled audio until we correctly configured separate AudioContext instances
  • Dual model pipeline: Coordinating Nova Sonic for conversation and Nova Lite for report generation required careful session management

What I learned

  • Deep understanding of Amazon Nova 2 Sonic's WebSocket streaming protocol
  • Real-time audio processing in the browser using AudioWorkletNode
  • How to build a complete clinical intake workflow from scratch
  • The importance of patient-side vs doctor-side healthcare AI solutions

What's next for MediVoice

  • EHR Integration — Connect with Epic and Athena for seamless doctor workflow
  • SMS Access — Send patients a link before appointment, no login required
  • 50+ Languages — Expand multilingual support using Nova Sonic's capabilities
  • ICD-10 Suggestions — Auto-suggest diagnosis codes from intake data
  • Wearable Integration — Pull in Apple Watch and Fitbit vitals automatically
  • Ambient Mode — Continue listening during the appointment for complete notes

🏆 Why MediVoice is different

Unlike Nuance DAX, Suki AI, and Epic's ambient documentation — which all assist doctors DURING appointments — MediVoice works BEFORE the appointment on the patient side. Zero workflow change for doctors. Immediate value from day one.

AmazonNova

Built With

Share this project:

Updates