Arya-Hackaton

architure diagram
RAG arch for context
page 1
motivation
dashboard with live transcription
database supabase
data model relational

Inspiration Healthcare intake is broken. Patients wait on hold for 20–40 minutes just to give their name and date of birth. Staff spend hours collecting data instead of delivering care. There's no triage — emergencies are treated the same as routine calls. We wanted to fix that with a single phone call.

What it does Arya Health Voice AI is a production-ready conversational agent that handles full patient intake over a phone call — from hello to care match. A patient dials in, the AI answers instantly, collects their name, date of birth, reason for visit, symptoms, allergies, medications, and conditions — then saves everything to a live database. A real-time dashboard shows the intake populating as the call happens.

Key features:

Patient recognition — returning patients get a personalized greeting Emergency detection — chest pain, stroke signs, or crisis keywords immediately redirect to 911 or 988 Full transcript logging — every turn of the conversation saved in real time Live dashboard — care coordinators see intake data appear as the call progresses How we built it Twilio — handles inbound voice calls, speech-to-text via , text-to-speech via Amazon Polly Google ADK — orchestrates the multi-step intake conversation with function-calling tools Groq + LLaMA 3.3 70B — powers the LLM responses (fast, ~0.2s per turn) Supabase — stores patients, calls, transcripts, and agent sessions with Realtime subscriptions FastAPI — webhook layer between Twilio and the agent Live Dashboard — pure HTML + Supabase Realtime, no framework needed Challenges we faced Small local models hallucinate values they haven't been told — we fixed this by using separate tools per field (save_name, save_date_of_birth) so the model can't fabricate a DOB on the name turn Twilio signature validation breaks behind ngrok tunnels — we reconstructed the original public URL from X-Forwarded-* headers Merging two parallel branches (voice webhook + knowledge base) with 27 conflicting commits under time pressure What we learned Voice AI for healthcare requires a different bar than chatbots — reliability, safety guardrails, and graceful degradation aren't extras, they're the prerequisite.