Healthbook

Inspiration

Our inspiration is personal. A lot of our family members are dealing with ongoing health problems, and we’ve seen how exhausting it can be to manage appointments on top of everything else. We wanted to reduce the hassle of scheduling necessary care so people can focus on their health, not on figuring out where to get help. Healthbook is our way of turning that frustration into something useful, accessible, and supportive.

What it does

Healthbook is an AI-powered medical booking assistant that helps users:

Create accounts and securely log in
Save personal preferences (insurance, location, timing, and provider preferences)
Search for doctors in their area based on those preferences
Use speech-to-text for voice input
Hear audio feedback and spoken summaries during each step of the process
Track and manage appointments in one dashboard

Healthbook is connected to Supabase for persistent memory, so users, preferences, and appointments are maintained across sessions.

How we built it

We built Healthbook as a full-stack system with real-time AI orchestration:

Frontend framework: Next.js + TypeScript + Tailwind CSS
Backend: Python (FastAPI sidecar) + Next.js API routes
Automation: Browser Use to navigate provider websites and search in real time
Data + Memory: Supabase (Postgres) + pgvector for persistent memory
Voice accessibility: Whisper (speech-to-text) + OpenAI(audio-output)
Calling/testing: Twilio + ngrok to test and validate live call flows
Realtime updates: SSE streaming to show live agent/search progress in the UI

We split work across frontend UX, backend/automation, and voice integrations, then merged and tested together end-to-end.

Challenges we ran into

One of our biggest challenges was runtime. Browser Use is powerful for web automation, but it does not provide persistent memory between sessions. Early on, the agent revisited the same doctor websites repeatedly, which slowed searches and made the flow less efficient.

We solved this by adding a vector-memory layer (Supabase + pgvector) to store and retrieve relevant prior context. That allowed our system to avoid duplicate revisits and focus on new, useful providers.

Other challenges included:

Managing authentication/session edge cases across routes
Coordinating multiple APIs/services under hackathon time pressure
Keeping branch merges stable while all three teammates were shipping in parallel

Accomplishments that we’re proud of

Built a true end-to-end healthcare assistant, not just a UI prototype
Made the platform significantly more accessible with speech-to-text and audio indicators throughout the user flow
Integrated Browser Use with persistent vector memory for practical automation
Reduced end-to-end runtime by approximately 97% by preventing duplicate revisits
Successfully used a Python backend and ngrok-based call testing to validate real communication workflows