An AI drive-thru assistant, built in under a day

"For the Big Mac can I have some fruit and white coffee. But I changed my mind, I want the chicken one large. Can I have water and fries with that? No pickles on the Big Mac, extra mayo on the chicken. Actually, make that a black coffee for the first menu."

Every drive-thru worker has heard an order like that. Most software would fall over. Mine untangled it and read the order back with a total, no-pickles, extra mayo and coffee swap included.

Why I built it

The big fast-food chains have spent tens of millions on AI ordering. I wanted to know how close one independent developer could get with GPT-4, the then-new function calling API and a few hours of focused work.

Answer: close enough to be interesting.

How it works

Personality from the staff handbook. I pulled the rules, responsibilities and tone straight out of a real new-employee manual. Friendly, efficient, never pushy. (I skipped the section on uniforms.)

Grounded in the real menu. Every item, price, option and meal size went in, along with opening hours, location and breakfast times. So when someone asks for a breakfast muffin at 12:13, the assistant politely explains breakfast finished at 11.

Orders as data. Function calling defines exactly what an order looks like: items, sizes, sides, drinks, modifications. The model keeps the conversation human while filling in that structure behind the scenes. What comes out the other end is clean JSON a kitchen screen or POS could consume directly.

It handled upsells, totals, "actually, remove that," dietary tweaks and questions about what's in a burger. It was ready to test in less than a day, because it was built on my React AI assistant starter kit, which already had the chat interface and model connection in place.

Beyond burgers

The same recipe fits anywhere people order, book or configure something by talking:

• Restaurants and cafés: drive-thru, kiosk, phone or table ordering
• Retail: guided product finders that turn a vague request into a basket
• Hospitality: room service, bookings and upgrades
• Field services: turning a customer's rambling description into a structured job ticket

The pattern is always the same. Brand voice and rules, grounded data, a strict output schema. The hard part is making it reliable at volume, with sensible fallbacks to a human, and that's the part I focus on. Let's talk about your use case.

More on this project

• The demo on YouTube
• The full build tutorial on YouTube (under 30 minutes)
• The full write-up on Medium