Who makes Zogu and how we test it
Zogu answers a restaurant's ordering line and takes the order from the menu in its POS. Ardit Mula builds it. As of 5 November 2026 it has only answered our own test calls, so there are no customer quotes here.
Who builds Zogu
[Photo: Ardit Mula]
Ardit Mula, [job title]
Ardit wrote Zogu's plan and sets the rules the phone agent follows. [One or two sentences from Ardit on his background, in his own words.]
Most of the code is written by an AI tool, Claude, working to those rules. It can put changes on the test system. Only Ardit puts anything live, and nothing is live yet.
[Anyone else on the team: name, role, a real photo and one line each, only for people who have agreed to be named.]
Photos here are real. We don't use AI-made faces.
Zogu and your POS
Zogu connects to restaurants' POS systems through our integration platform, and is registered on its test system only. It sells only what's on a restaurant's POS menu, at its prices, and sends orders back into that POS.
[Our relationship with our integration platform, in the exact words agreed with them.]
Why we're building it
[Ardit, in your own words: why you started Zogu. Two or three sentences.]
The plan's pilot targets are a 0.8-second median reply and 95% of orders right. On test calls the median is 1.3 seconds, and accuracy hasn't been measured yet. The rules:
- Every call opens by saying it's the automated ordering assistant.
- No transfers or promised call-backs.
- One read-back and a clear yes, or nothing is sent.
- Zogu's code sets prices and totals, never the AI model.
- No card details over the phone, and no waiting for the kitchen.
More on how Zogu takes a phone order, step by step and what Zogu doesn't do.
How we test it on real phone calls
We ring a pickup-only test restaurant on our integration platform's test system, on a US test number. Of 20 test calls on 4 and 5 November 2026:
- 7 ended with an order, all accepted or completed.
- 6 ended without one, some on purpose.
- 4 failed early, while we fixed first bugs and after our free Cloudflare AI allowance ran out.
- 2 were abandoned, and 1 ended with an apology.
We made those calls ourselves, so it isn't a success rate. An AI tool does the first pass on each call record; Ardit decides what to fix. Every fix gets a test that fails on the old code.
What the test calls found
- "With cheese" after three doner kebabs got the cheesecake. Now: "Sorry, the Doner Kebabs don't come with a cheese option."
- A misheard "and a doner kebab" was saved as a name. Now it asks again (the details).
- The model offered a transfer. The ban is now in code, not only the instructions.
- One caller heard 12 confirm-style questions. Replayed through today's build, it's 3.
How this website was written
Claude, made by Anthropic, drafted this site from our product notes, code and call records. Every claim had to point to the code or test behind it. A person then checks each page. [Name] checked this page on [date].
The phone agent itself doesn't use Claude: it runs on Cloudflare Workers AI. If something here is wrong, tell us through the contact page.
Company details
[Company name] is registered in [country of registration], company number [Company number]. Registered office: [Registered address].
Questions and answers
Is any restaurant using Zogu yet?
No. It needs production phone numbers, a real POS install and a live console first, and none of that is in place.
Does the phone agent run on Claude?
No. It uses Deepgram Flux, GLM-4.7-Flash and Deepgram Aura-2 on Cloudflare Workers AI, with calls over Twilio.
Are calls recorded, and who can see them?
No audio. Text transcripts sit in our internal console, behind a Google sign-in. See AI disclosure and call transcripts.