By Juste Thales Gnimavo (CEO, ZeroSuite) and Claude Opus 5.5 (Claude Code)
You speak French. The person in front of you speaks English. Each of you talks in your own language and hears the other in yours, live, without pausing after every sentence. When the call ends, you get a written report of what was said and agreed, numbers included.
That is Déblo Duo. It has been running in production on the web since late October 2026, and the native app for iPhone and Android was built on November 2. This is the first article of a series on how it was made. It starts with the decision that shaped everything else: Duo is a separate app, and we wrote down why, so that it does not become a habit.
It started as a drawer
Duo did not begin as a product. It began as a feature inside Déblo, the voice-first assistant we build from Abidjan. On October 20 we ran a first test of a real-time speech translation engine on a simulated negotiation in French and English. On October 21 we tested it on real voices and placed real calls in production: a three-minute call billed for exactly 186 seconds, and a remote call with someone abroad in which each side heard only the other's translation.
The plan at that point was modest. The interpreter would live in the Déblo app's menu drawer, next to everything else. Phase four of the brief said so.
On October 21, Juste changed it. The interpreter would be a product of its own.
The next morning, when Claude opened the framing session, the brief contradicted itself. The decision line said "separate product". The progress section, the ambition paragraph and the note pointing to the next prompt still prescribed the drawer. A session started from that note would have built a decision that no longer existed. The first thing the framing session did was fix the brief. The second was to refuse to write code: Juste had asked for a discussion only, and that is what it stayed.
One app or two
A second app is not free. Every app we ship costs a store listing, a review cycle on two stores, its own builds, and one more install we ask a stranger to make. Splitting a product into many small apps is also how a suite becomes a pile.
So before agreeing, Claude looked at how others handle it. The most useful reference was closest to home: Anthropic's own App Store listings. There were two apps. One is the full Claude app. The other is the same app packaged for companies that manage phones through Microsoft Intune. Anthropic does not split by feature. The second app exists for a distribution constraint, not because a feature wanted its own icon. (Claude Duet, announced on October 21, the day before this session, is a proactive agent, not an interpreter, and was not on the store yet.)
That gave us the rule. A module leaves the Déblo app and becomes its own app only when two things are true:
- it serves a distinct moment of use, one that is different from opening an assistant;
- its measured usage inside the main app is above a threshold.
Duo meets the first condition plainly. You open it in a few seconds, standing in front of someone you cannot understand, before a call abroad, or before stepping onto a stage. That moment has nothing to do with asking an assistant a question. It fails the second condition, because it had no measured usage yet. We said so in writing: Duo is the exception, released before measurement, because the moment is obvious and because it is the product we show when we pitch. The threshold for the next module is still to be set, and that line is in the log as an open item, not a forgotten one.
Same account, same balance
Separate app does not mean separate company. The decisions Juste took that day keep Duo attached to everything Déblo already has:
- the same backend and the same account, signed in with the same code on the web, in the Déblo app and in the Duo app;
- the same balance, topped up once, spent anywhere;
- no separate website: Duo is a tool of deblo.ai, at
/duo, with a public page that sends each visitor to the right store; - a "Déblo Duo" entry in the Déblo app's menu that opens the Duo app.
The Déblo app stays the base. Other modules may leave it later, if the market asks for them and if they pass the rule. Voice is the obvious candidate. Nothing else is scheduled.
The guest pays nothing
The product decision we care about most is about the other person. The guest joins through a QR code or a short link, from a browser. No app, no account, nothing to pay. The host chooses whether to protect the link with a short PIN, which they can read out over the phone, and how long it stays valid.
We also decided that the guest picks their own language. The host often does not know it. Think of the taxi driver, the hotel receptionist, or the landlord of the flat you just rented. You cannot ask them to fill in a form before you can talk.
There is a reason beyond courtesy. The production numbers we read that day showed that voice carries Déblo's usage, and that guests without accounts make up most of it. Every Duo call puts someone who has never heard of Déblo in front of it, during a real conversation and not an advertisement. At the end of the call, their page offers them the app. In the "I'm presenting" mode, a whole room uses it at once. Acquisition is part of the usage itself, and that only works if the guest's first step costs them nothing.
A name already taken three times
"Duo" is a good word for a two-person conversation, and a crowded one. Google had a calling app called Duo. Claude Duet had launched the day before. And in the language-learning category, Duo is the name of Duolingo's mascot. There is even an existing app called "Duo Interpreter".
So "Déblo" always stays attached to "Duo", and we kept a fallback name ready ("Déblo Link") until the brand check came back. On October 23 the store names were approved: "Déblo Duo: Live AI Interpreter" in English, "Déblo Duo : Interprète IA" in French.
Interpreter, not translator
The choice of "interpreter" over "translator" came from the stores, not from taste. We searched both. The "voice translator" category is saturated: in most of the countries we checked, the top positions belong to apps with tens of thousands of ratings, and they all describe themselves the same way, with text, photo and voice. "Interpreter" is a word people search for, and the autocomplete suggests it, yet on October 23 the best dedicated app on the US App Store had 40 ratings.
The name takes the place nobody holds. "Translator" is still the word most people type, so it goes in the subtitle, where the store indexes it too.
We do not announce languages we have not heard
One more decision from the framing session deserves to be written down, because it is the kind that slips. The underlying engine supports many languages. A launch page saying "70 languages" would have been easy to write. We refused it. A language is announced once it has passed a test on real voices, not once a setting allows it. Eleven languages were open in production at the end of October, opened and closed from the admin without an app release. The public number follows the proven ones.
What changed since
On November 1, Duo gained a second moment, on equal footing with the interpreter: Notes. You dictate up to two hours, even switching between languages, and get back a transcript, a summary with every figure copied exactly, and both translated into the language you choose to read them in. Plain transcription is free on every phone now, so Notes is not sold on transcription. It is sold on the summary, and on reading in a different language from the one you recorded in. Juste uses it for his own voice notes.
Notes does not break the rule. It is the same app and the same moment of "I have something to keep in a language", not a third icon.
The next articles go into how Duo was built: the first test on Juste's own voice, packets that disappeared without an error, a conference room that started translating itself, the free trial we refused before building it, and why a product made in Abidjan lists Africa last among its markets.
Duo is on the web today at deblo.ai/duo. The store release is next.