Case studyFrontDesk

An AI front desk that books the appointment.

FrontDesk is an AI receptionist for salons, med spas and clinics. It answers Instagram DMs and phone calls in French or English, books into the business's own Google Calendar, and hands off to a person when a request needs one.

Status
Live. Offered as free pilot projects to a few Montréal businesses.
Role
Scoped, architected and shipped solo by Frederic.
Built with
  • Next.js
  • TypeScript
  • Vercel AI SDK
  • OpenRouter
  • Vapi
  • Twilio
  • Neon Postgres
  • Drizzle
  • Clerk
  • Google Calendar API

The problem

A booking request rarely arrives at a good time: a DM in the evening, a call while everyone is with a client. FrontDesk answers the moment the message or the call comes in, and it only replies to people who wrote or called first.

What was built

  • One agent per business, configured from a versioned setup: services, hours, staff, FAQ and guardrails that can be switched on or off.
  • Instagram DMs and inbound phone calls, answered from the same configuration, in French or English.
  • Booking straight into the business's Google Calendar, which stays the source of truth. If the calendar does not answer, the agent says so instead of guessing a slot.
  • Hand-off to a person: the owner gets an email, the agent stays paused until the conversation is handed back, and calls can transfer to a phone, by default only during opening hours.
  • An invitation-only inbox for each business, where staff can reply and hand the conversation back.
  • An operator console with the conversations and metrics such as the escalation rate.

How it works

  • One Next.js app on Vercel, with Postgres on Neon through Drizzle.
  • Models are reached through OpenRouter, so each business's model is a setting, and every request opts out of provider data collection.
  • The voice assistant is rendered for each call from the same live configuration as the DM agent, so the two channels cannot drift apart.
  • Incoming messages are stored in a Postgres outbox and processed in order, with retries and a scheduled sweep, so a crash does not lose one and no queue service is needed.
  • Calendar events carry only a first name and last initial, the service, the practitioner and a note that the assistant booked it.
  • An evaluation suite of 13 scenarios, run on Instagram and on voice where they apply, is scored by a language-model judge.