Resse.ai — the front desk that answers itself

Small businesses lose customers to a phone nobody is free to answer.

The receptionist is busy with the person in front of them, so the person walking in waits, and the person calling gives up. Resse.ai takes the walk-in half: it greets, answers what the business actually offers, finds a real open slot, and holds the booking until a human confirms it.

A real exchange, from the tablet above

Nothing here is scripted. This is what the agent said when asked for a service the clinic doesn't list.

Visitor

“I would like to book a general consultation.”

Resse.ai

General consultation isn’t one of the listed services. Would you like me to book it as a Checkup at ten in the morning on Monday, 14 September?

It didn’t invent a service to be agreeable, and it didn’t invent a time. The slot came from the clinic’s opening hours minus everything already booked — including the owner’s Google Calendar. Confirming it is a human’s tap, not the agent’s decision.

What happens when someone walks up

It notices

No wake word, no button

Face detection runs on the device itself — MediaPipe compiled to WebAssembly, inside a hidden WebView — so it costs nothing per check and no camera frame ever leaves the tablet.

It listens

And knows when you’ve finished

Recording stops on silence rather than on a timer, so short answers aren’t padded and long ones aren’t cut off.

It answers

Out loud, with the avatar’s mouth moving and captions on screen

Then it listens again, immediately, for the follow-up — the thing that separates a conversation from a search box.

It books

Into a real calendar, against a real deposit

Confirming writes a Google Calendar event. If the business takes a deposit, a UPI QR appears on the approval card first.

The owner dashboard on a tablet, showing a clinic's hours, services as chips, description, and booking deposit, imported from the clinic's website. — Set up by pasting a URL. The services, description and hours here were read off the clinic’s own website.

The owner gets the other half

Point it at a business’s website and it builds the receptionist’s knowledge from what it finds. Behind that sits an admin surface: today’s numbers, the client directory, an appointments calendar, and settings for the voice, the avatar and the integrations.

Nothing is written without a tap. Reads render cards; writes render an approval card.
It never dials or texts on its own. It opens the phone’s own dialer or composer with the fields filled in.
It knows what day it is. The kiosk’s real date and timezone go to the model every turn, so “next Tuesday” resolves to a date that exists.

How it’s put together

An Expo app holds the tools and the app state; a Next.js backend holds the CopilotKit runtime, the API keys, and the speech routes. Tool calls execute on the device — that’s how the agent can open a dialer, render a QR, or block on someone’s approval tap — while the model runs server-side. Swapping the model provider, or the whole agent framework, is a change in one file.

What it isn’t yet

Payment confirmation is a person looking at the payer’s phone. There’s no gateway webhook.
The Google session lasts about an hour; there’s no refresh flow.
Appointments live on the device. No server-side store, no multi-device sync.
Rescheduling and cancelling an existing booking aren’t built.

Read the code on GitHub Watch the demo

Built in 4 Hours

Resse.ai was made by Balaji Anbalagan and Saisathish Karthikeyan for AI Tinkerers’ Agents, Everywhere hackathon in September 2026, starting from CopilotKit’s starter kit.

The constraint that shaped most of it: stay inside Expo Go, so the whole thing runs by scanning a QR code instead of producing a build. That ruled out every native face-detection module — which is why detection ended up as WebAssembly in a hidden WebView.

Demo video Post on X Post on LinkedIn

Balaji and Saisathish at a desk with two laptops open, building Resse.ai during the hackathon. — Balaji and Saisathish, mid-build.