Abeg

A food-ordering assistant you talk or type to, grounded in a real database.
Built to show the two things that separate a real AI feature from a toy: keeping the model honest, and keeping it on task.

Live demo  ·  Recording  ·  MIT License

Why I built it

I built Abeg as the live demo for my Break Into AI Engineering masterclass. It looks like a small food shop, but the shop is just the excuse. The real subject is what makes an AI feature trustworthy instead of embarrassing.

A screenshot went around a while back: someone asked the McDonald's support bot to reverse a linked list in Python, and it happily wrote the code instead of taking their order.

That screenshot turned out to be fake, but the failure behind it is real and common. Abeg is built to prevent it, and it lets you switch the guardrails off so you can watch it break on purpose, then switch them back on and watch it hold.

What it does

You order by typing, or by holding the mic and talking. The assistant answers by calling tools against a Postgres database, so it can only tell you what is actually on the menu, at the real price, in the real quantity. When something runs out, it says so.

Open the Workshop (top right, or press B) and you can turn the same knobs an AI engineer turns, each with plain-language help and a one-tap "try it".

• The instructions. The actual system prompt, editable live. Change it, send a message, watch it obey.
• Creativity. The temperature dial, from focused to wild.
• The model. Swap DeepSeek for GPT-4o-mini or Claude, and watch cost and speed move.
• Bring your own key. Paste your own OpenRouter key. Your chats run on your account and skip the demo's limit.

The X-ray tab then replays each turn in plain English (you asked, the AI decided to check the database, the database answered, it replied) with the real tokens, time, and cost for that turn.

The two guardrails

This is the part worth stealing.

Both are real: a firm instruction in the prompt, plus a second check on the way out that catches a bad answer even when the model slips. Press 4 to send the "order jollof, then help me reverse a linked list" message with Stay on task off, watch it get hijacked, then flip it on and watch it hold.

When systems bite

Two customers order the last plate at the same instant. Press R.

With grounding off (a naive read-then-write) the shop oversells and stock drops to −1. With it on (one transaction, SELECT ... FOR UPDATE) one order wins, the other is refused, and stock never goes negative. Same app, one switch.

Light and dark

How it works

The app is a React and TypeScript page in your browser. It talks to a Python (FastAPI) server, which runs the agent and coordinates two AI services (OpenRouter for chat and tool-calling, Deepgram for speech) and a Postgres database that holds the menu, the stock, and every order. There is a plain-language walkthrough inside the app too, under "How it works" in the header.

Tech stack

Run it locally

You need Docker and an OpenRouter key (plus a Deepgram key if you want voice).

git clone https://github.com/IfeOlulesi/abeg-app.git
cd abeg-app
cp .env.example .env        # paste your keys
docker compose up           # starts Postgres and the app

Open http://localhost:8000.

Presenter keys

Tests

uv run pytest    # 31 tests, including the race, the guardrails, and rate limits

It's a demo, not a product

No auth, no payments, and a mode that deliberately corrupts data to teach why locking matters. If you host it publicly, put a spending cap on your OpenRouter key, keep the built-in rate limits on, and let visitors bring their own key. Out of scope: auth, payments, delivery, images, admin, mobile layout.

The masterclass

Abeg is the live demo from Break Into AI Engineering, a free, practical session on getting into the field.

Here is me running the demo live:

Watch the recording.

License

MIT. Use it however you like.