Back to projects

Voice AI Restaurant Ordering Prototype

Built as a graded prototype around a fictional Nigerian restaurant scenario. It handled test calls and sample data, but has not yet been validated in live restaurant operations.

Status
Academy prototype
Category
Voice AI
Tools
Vapi / n8n / Webhooks / Google Sheets / REST APIs
View repository
Workflow diagram for the restaurant voice ordering prototype
Test path from Vapi through n8n to Google Sheets

Results and next steps

Result

A test call reached n8n and routed sample order and reservation fields into Google Sheets.

Current boundary

Dish-name transcription, interruptions, ambiguous quantities, and missing fields need stronger validation.

Next improvement

Add field-level confirmation and test noisy calls, interruptions, and incomplete orders before any live pilot.

How information moves

  1. EnquiryA test caller describes an order or reservation.
  2. CaptureVapi transcribes the call and collects the details.
  3. Routen8n receives the webhook and applies the workflow rules.
  4. RecordThe sample fields are written to Google Sheets.
  5. ReviewA person checks missing or uncertain information.

Overview

This was my TS Academy final project. I built it around a fictional Nigerian restaurant scenario so I could test a complete voice ordering workflow without presenting the work as a live client deployment.

The problem

A restaurant can lose time when staff repeatedly answer routine calls, copy order details, and correct missing information. The prototype explored whether a voice agent could collect structured order or reservation details and pass them into a workflow for review.

How the system works

  1. A caller describes an order or reservation.
  2. Vapi captures and transcribes the request.
  3. A webhook sends the structured call data to n8n.
  4. n8n applies the workflow rules and writes sample fields to Google Sheets.
  5. A person reviews missing or uncertain details.

What I tested

A test call reached n8n and routed sample order and reservation fields into Google Sheets. This demonstrated the end-to-end path with sample data.

Known limits

The prototype still needs stronger handling for dish-name transcription, callers interrupting the agent, ambiguous quantities, noisy calls, and missing required fields. It has not been validated in live restaurant operations.

What I would improve next

I would add field-level confirmation, structured retries, clearer human handoff rules, and a test matrix covering incomplete orders and transcription failures before proposing a live pilot.