Mini Eats
Let's run a restaurant's lunch rush: every order updates the kitchen's stock, and the AI never orders on its own.
The story
Mini Eats gives a small restaurant live, recipe-based inventory: every sale deducts its ingredients, and the manager sees what is running low or about to expire. Customers order ahead from a QR code at the restaurant, the website or an iPhone app prototype. A small team built it in a three-day course with Anthropic and CodePath at Miami Dade College, for a real restaurant-tech company.
One rule runs through it: plain code computes every number, and the AI only suggests. The restaurant and its data are invented.
My role
I built the frontend of the customer ordering app (voice ordering, group orders, AI suggestions, English and Spanish) and worked on the manager login. Teammates built the inventory backend, the manager dashboard, the iOS apps and the first prototype. Team project from an intensive course, built with AI coding tools.
Overview
Mini Eats is a restaurant ordering and inventory platform built by a small team in a three-day course (AI Elite Basecamp, with Anthropic and CodePath, at Miami Dade College) for a real restaurant-tech company. Its core is recipe-based inventory: every sale deducts ingredients, stock is tracked by batch, and the manager gets low-stock and expiry alerts. Customers order from a QR code, the website or an iPhone app prototype. All data is synthetic, and nobody pays in the app.
My role. I built the frontend of the customer web app: QR ordering, English and Spanish, order tracking, allergen badges, AI suggestions, voice ordering and group orders. I also worked on the manager login. Teammates built the backend, the manager dashboard, the iOS apps and the first prototype; credits are in the README. I built my part with Claude, deciding what to build and testing every change.
Architecture
- Backend (Express, Prisma, PostgreSQL): the single source of truth for menu, recipes, inventory, orders, alerts and sales.
- Manager dashboard (React, Vite): live stock and servings left per dish, alerts with one-tap actions (reorder, take a dish off the menu), an orders board, sales and an end-of-day summary.
- Customer web app (Next.js, React, Tailwind): calls the backend and keeps only its own data in Firestore. Suggestions and voice use one structured-output call to Claude or Gemini, and Zod validates every answer.
- iOS apps (SwiftUI, prototypes): a customer app and a merchant app on the same backend, so a phone order and a web order are the same ticket in the kitchen.
Key decisions
- Plain code computes every number. Stock, totals and alerts are code. The AI only writes language: suggestions, voice replies and the end-of-day summary.
- The AI proposes, code decides. For voice, the model returns dish ids and quantities. Code checks them against the live stock, drops anything sold out or on the avoid list, trims to the budget and computes the total.
- Voice never sends an order. Dishes go into the cart, the customer can remove any of them, and nothing reaches the kitchen until they confirm.
- Conversation memory. Each voice session remembers its last suggestions, so "order that" works.
- Group orders. A 4-letter code, up to 12 people, and a split in cents that adds up exactly.
Testing and evaluation
There is no automated test suite for the customer app, only a read-only end-to-end script (scripts/smoke.mjs), because a real order deducts real inventory. I tested each change by hand, and the team tested the full flow together. These were manual checks, so there are no accuracy numbers.
What I'd do next
- Turn the manual voice and group-order checks into automated tests.
- Build a voice eval set: misheard words, ambiguous dishes, allergies and "that one" follow-ups.
- Next.js
- React
- TypeScript
- Tailwind CSS
- Firebase Firestore
- Claude API
- Gemini API
- Web Speech API
- Zod
- Express
- Prisma
- PostgreSQL
- SwiftUI