ClinicFlow

🩺 About the Project — ClinicFlow

Inspiration

Clinical documentation is one of the biggest sources of burnout for healthcare providers. Doctors often spend more time typing notes than engaging with patients. We wanted to explore whether a voice-first, AI-assisted workflow could reduce this burden without disrupting how clinicians already think and speak during a visit.

ClinicFlow was inspired by the idea that doctors naturally reason out loud, and that voice notes contain rich clinical context that is often lost when translated into rigid forms.

What We Built

ClinicFlow allows clinicians to:

Record a short voice note after a patient encounter
Automatically transcribe the audio using speech-to-text
Convert the transcript into a structured SOAP note using LLMs
Save the result as a clinical visit that can be:
- reviewed visually
- navigated in a visit list
- played back as an audio summary

The system supports:

Visit list and visit detail views
Structured Subjective, Objective, Assessment, and Plan sections
Voice playback of visit summaries
A modular backend designed to plug in different AI agents and services

How We Built It

Frontend

React + Vite for fast iteration
Tailwind for clean, responsive layouts
Browser MediaRecorder API for voice capture
REST-based integration with the backend

Backend

FastAPI for API routing
OpenAI Speech-to-Text for transcription
LLM-based agent pipeline to transform transcripts into SOAP notes
OpenAI Text-to-Speech for audio visit summaries
In-memory visit storage for rapid prototyping

The architecture was designed to be service-oriented, allowing speech, reasoning, and synthesis to evolve independently.

Challenges We Faced

Handling audio capture and uploads reliably across browsers
Managing async workflows from voice → transcript → structured data
Designing prompts that generate clinically reasonable SOAP notes
Avoiding over-automation while still providing meaningful structure
Dealing with local development issues around file systems and tooling

Each challenge pushed us to simplify and focus on reliability over complexity.

What We Learned

Voice is a powerful interface for clinical workflows when paired with structure
Small, focused AI agents are more controllable than monolithic prompts
Clear separation between UI, transcription, reasoning, and synthesis is critical
Developer experience matters even in healthcare prototypes

Most importantly, we learned that AI works best when it supports how humans already think, rather than forcing new behaviors.