An AI that answers a restaurant's phone and takes the order

The phone rings most when the kitchen is busiest, and every call taken at the counter pulls someone away from the food. The restaurant's online ordering system already holds the menu, the prices and the kitchen screen, so the goal is a phone agent that takes orders through that same system and can never quote a price the website wouldn't.

1. 01Not live

   Answers and takes the order

   A friendly voice greets the caller, asks pickup or delivery, takes items one at a time, reads back the total, confirms, and sends the order with an order code and a ready time.

2. 02Not live

   Talks like a person

   It listens while it speaks, so callers can interrupt the way they would with a person, and one server handles many calls at once. Speech-to-text, the language model and the voice run as one live stream.

3. 03Not live

   Works on a normal phone number

   Calls come in on a regular phone line and are passed into a call room the agent joins. The room closes when the caller hangs up.

4. 04Not live

   Uses the real menu every time

   The agent has 14 ordering tools: search the menu, get options and extras by size, price an item, look up a returning caller, set pickup or delivery, check an address, add or remove items, read back the cart, quote the total and submit.

5. 05Not live

   Knows returning callers

   It looks up the caller's number, greets a repeat customer by name and offers their usual delivery address.

6. 06Not live

   Exact prices and address checks

   Prices come from the ordering system, not from the AI's own math. A spoken address is looked up and measured by driving distance before any delivery fee is added.

7. 07Not live

   House rules on every call

   It never takes card numbers over the phone, never makes up an item or a price, explains problems in plain words, offers pickup when an address is too far, and handles a closed kitchen politely.

8. 08Not live

   A fresh cart for each call

   Each call has its own cart, saved to the database once when the order is placed and wiped on hang-up, so the next caller never inherits it.

Node.js

Runs the agent on a server.

TypeScript

JavaScript with types, so many mistakes are caught before the code runs.

LiveKit Agents

Connects a phone call to the AI and handles who is talking when.

Deepgram Nova-3

Speech-to-text: turns what the caller says into text as they speak.

OpenAI gpt-4o-mini

The language model that follows the conversation and calls the ordering tools.

Cartesia Sonic Turbo

Text-to-speech: turns the agent's replies into a natural voice with very little delay.

Silero VAD

Voice activity detection: notices when the caller starts and stops talking.

Zod

Checks every piece of incoming data before it is used.

The client isn't named and nothing here comes from their data. The animation is a concept with sample data.