01
Booking by phone
Vapi runs the call and invokes tools that are Convex HTTP actions. They check Google Calendar, then book, reschedule or cancel while the patient is still on the line.
Case study 02
My own productPre-launch
DentalAI Receptionist
An AI receptionist that answers a dental clinic's phone, books the appointment and knows when to hand over to a person. Clinic staff watch every call from a multi-tenant dashboard.
- Role
- Product, design, backend and voice flow
- Timeline
- Aug 2026 to now
- Stack
- Vapi, Convex, Next.js, Google Calendar, n8n, Claude API
- Scale
- Multi-tenant from day one
The problem
A small dental clinic misses calls while the front desk is busy, and every missed call can be a missed booking. The receptionist has to pick up every time, book correctly and never pretend to be a clinician.
The hard requirements: book against the clinic's real calendar during the call, and hand anything clinical, urgent or about money to a person.
What I built
02
Knowing when to hand over
Emergencies, billing disputes and clinical questions go to staff. The AI's only moves on a clinical question are to answer from the clinic's FAQ or escalate.
03
A dashboard for the clinic
Owners and staff see calls, appointments, the knowledge base and AI settings. After each call the Claude API writes a summary, and n8n sends the patient an SMS confirmation.
How it fits together
The caller talks to Vapi, Vapi calls Convex, and Convex is the whole backend.
- Callerpatient on the phone
- Vapinumber, speech, turn-taking
- Convex HTTP actionstools, webhooks, tenant data
- Google Calendarreal availability
- Claude APIcall summaries
- n8nSMS confirmations
Decisions and trade-offs
Vapi instead of a home-made voice stack
Telephony, speech-to-text, barge-in and text-to-speech latency are hard real-time problems. Vapi owns them, and my code owns the dental rules.
Trade-off: per-minute cost and less control over the speech pipeline.
Convex as the entire backend
Queries, mutations, actions and HTTP actions are the backend, so there is no second service to deploy, and the dashboard gets live call lists for free.
Trade-off: no row-level security, so every function enforces the clinic boundary in code.
A hard line on what the AI won't do
No diagnosis, insurance claims or payments. Anything that needs clinical judgement is escalated to a person.
Trade-off: more calls reach staff than a looser assistant would send.
Results
- 6
- services behind one phone number
- 1
- backend deployment: Convex
- 3
- kinds of call always handed to a person