MDRhythm Voice Agent

One area where I excel as an AI PM: building agentic workflows from scratch, architecting the state machine, writing the prompts, and running a real clinic trial.

The Feature

A real-time voice scheduling agent trialled at a small GI clinic in the Bronx. Patients call in, get identified by phone number, then either book an appointment or leave a note — no staff needed for routine calls.

How I Built It

I designed the LangGraph state machine, wrote the GPT-4o-mini intent prompts, wired the Deepgram WebSocket pipeline, and debugged audio encoding at the binary level. PM and builder on the same ticket.

Why It Matters

AI features fail at the seams — between STT and intent, between intent and action. Being in the code means I catch those gaps before they become missed appointments or patient-facing errors.

Orchestration Flow

Booking path
Note path / New patient
Intent routing
Animated = live data flow
BookingNoteReturningNew
📞 Patient Calls In
🎙️ Collect Mobile Number
register_mobile()
🧠 Detect Intent
set_intent()
⬡ Booking or Note?
👤 New or Returning?
set_patient_type()
🔁 Returning — Verify DOB
fetch_patient()
✨ New — Spell Name + DOB
register_new_patient()
📅 Check Availability
check_availability()
✅ Confirm & Book
book_appointment()
🏁 Goodbye
✍️ Spell Name Letter by Letter
name confirmed
📝 Collect Note Content
leave_note()
🏁 Goodbye
Mini Map

Tap any node to learn what it does