How I built Aria: an inventory app with an AI voice assistant

Anyone who works in a restaurant knows the stockroom never stops. Deliveries arrive, supplier prices change, something runs out earlier than expected, and at the end of the month there's a stocktake to do by hand. At Scaravelli, a restaurant in Mantua, Italy, all of this lived across spreadsheets, supplier price lists and paper counts.

That's where Aria came from: a custom inventory app with a voice assistant you simply talk to. In this article I explain how I designed and built it, the choices I made and what I learned.

The problem: too many spreadsheets, too many prices to re-check

The starting point wasn't "let's put AI somewhere". It was much more concrete:

• product prices had to be re-checked every time, because price lists change;
• every stocktake took hours of counting and copying by hand;
• data was spread across several files, so it was hard to see the total stock value at a glance.

Off-the-shelf software would have solved part of the problem, but it would have forced the staff to change their habits. I chose to build a tool around the way they already work.

The catalogue: 5,300 products and 60 suppliers

The first step was getting the data in order. I built the catalogue from a dataset of about 5,300 products and 60 suppliers, with a price history for every item. When a price changes, the app knows and updates it automatically in stocktake sessions too.

On top of the catalogue there's a dashboard with items moved, stock value at purchase price and categories, plus an inventory table with search, filters, export and camera support.

The voice assistant: talking to the app

The most interesting part is Aria. In a kitchen or stockroom your hands are often busy, so typing on a tablet isn't convenient. With Aria you tap a button and speak, or type if you prefer.

You can ask Aria to add or remove products, check stock or search the catalogue. Behind the scenes:

• reasoning is handled by Claude, Anthropic's model, which understands the request and turns it into a precise action in the app;
• the voice runs on ElevenLabs, for natural, fast replies;
• the app actually performs the operation and shows the result, so staff always see what happened.

My rule: the AI must never make things up. It only works on the app's data, and every action stays visible and under control.

The tech stack

• React, Vite and TypeScript for the interface, with Tailwind and shadcn/ui for components;
• Node.js with Express for the backend and inventory logic;
• Claude API for the assistant's reasoning;
• ElevenLabs for the voice.

The result

Aria now runs the stock of 3 stores in the same group, each with its own inventory, from a single app. Switching between stores takes one click and prices always stay in sync.

What I learned

1. Start from the problem, not the technology. The voice assistant works because it solves a real annoyance (busy hands), not because it looks cool.
2. Data comes before AI. Without a clean catalogue and reliable price history, no assistant can give useful answers.
3. Custom software adapts to people, not the other way round. That's what makes the difference compared to generic tools.

If you run a business and recognise these problems, get in touch: I'll explain how it could work for you.