By Juste Thales Gnimavo (CEO, ZeroSuite) and Claude Opus 5.5 (Claude Code)
On November 6, 2026, at 05:27, Déblo Duo went live on Google Play as Déblo Duo: Live AI Interpreter, published by ZeroSuite, Inc., in 178 countries. It had been submitted the evening before, at 22:05 UTC. The iPhone version was submitted to Apple at the same time, with its in-app purchases, and is waiting for review.
Get Déblo Duo on Google Play · On the web: deblo.ai/duo · Presentation site: debloduo.com
I have been shipping products since 2005, under many names. At the end, the same feeling always comes back: something is still missing. Usually I am wrong, and the feeling is only friction. This time it was different. Duo is the app I have wanted to put on the market for years. If someone offered me half a million dollars tomorrow to build a different one, I would say no.
This post is about why, and about the argument we had over when it was ready.
What Duo does
You speak the language you think in. The other person hears theirs, live, on their own phone. Translation starts before you finish your sentence. When the call ends, you get a report: the subject, the figures exactly as they were spoken, what was agreed, the open points.
The guest installs nothing. No app, no account, no payment. They join from their phone's browser.
Since version 1.1.0, the one now on Google Play, the home screen opens on your own code. That is Scan and Talk. The person in front of you scans it with their camera. Your phone rings like an incoming call: green to accept, red to decline. You accept, and each of you talks in your own language. Nothing starts without your consent. An unanswered request lapses after 60 seconds, the code on screen changes every 20 minutes, and the person scanning sees only your first name, never your number.
Around that, Duo offers several ways to talk:
- One phone between two people. For the taxi or the hotel desk. The top half of the screen is turned toward the other person, in their language.
- Two phones in the same room, through the code or a QR invitation.
- Remotely, by link. A short link sent on WhatsApp or by email, with an optional PIN and an expiry. One link serves one guest.
- "I'm presenting." One speaker, up to 100 listeners in one to three languages, each listening on their own phone, with subtitles. Simultaneous interpretation without booths or headsets.
- A printed code for a counter. A sticker for a front desk, a stand or a shop. It stays valid until its owner generates a new one.
- Notes. Dictate up to two hours, even mixing languages. Duo transcribes, tells speakers apart, summarizes, translates into another language in one tap, and exports to PDF, Word, Markdown or plain text.
Live interpretation runs on Google's real-time speech translation model (Gemini Live Translate, on Vertex AI), with LiveKit carrying the audio. It is the same infrastructure that already runs Déblo's voice assistant in production. Conversations work in 11 languages, in any pair: French, English, Spanish, Portuguese (Portugal and Brazil), Arabic, Simplified Chinese, German, Korean, Japanese, Turkish and Hindi. The interface is available in 19 languages, with Arabic mirrored right to left.
Only the host pays: $0.29 per minute per target language, billed by the second, and only while both people are online. Waiting for the guest costs nothing. Notes cost $0.02 per recorded minute. There is no subscription. The host tops up a wallet, the same one as in Déblo, with the same account. A three-minute trial works without an account, and every new account gets a welcome bonus.
Why Duo is a separate app and not a feature of Déblo is the subject of an earlier post.
The argument
I build with Claude Code: several sessions in iTerm, all day. They draft, implement, audit, and at the end of each session they write down where the project stands. That discipline has its own post.
Duo moved fast. We framed it on October 22. The web version went into production that same day, as a beta on deblo.ai/duo. The first iPhone build reached TestFlight on October 27. Then came more than thirty builds, Notes on November 1, Scan and Talk and the printed code in version 1.1.0 on November 4, and the store submission on November 5.
In those two weeks, we argued. The agents kept telling me the same thing in different words: Juste, you keep pushing it back. We have this, this and this. It is already good. Ship it.
And I kept answering: No. Not yet.
From the outside, it looks like the classic founder disease: perfectionism that kills a product before it meets a single user. I know that disease. I have had it. On other projects, the agents were right, I was wrong, and I lost weeks I did not need to lose.
On Duo, I was not waiting for a feature. The numbers were already there. On October 20, in a simulated negotiation in French and English, translation ended on average less than three seconds after the speaker. On October 21, a two-and-a-half-minute team meeting, in French as spoken in Abidjan, was translated faithfully from start to finish. On an order dictated with seventeen amounts, dates and percentages, Duo heard sixteen correctly. And a real call of three minutes six seconds was billed exactly 186 seconds. By any checklist, it was ready.
What I was waiting for was the moment I would want to use it. Not test it. Use it, without thinking about it.
The moment it was finished
I knew it was finished when it became my daily tool before it was even public.
I dictate the notes for my videos in French and read them back in English. I listen to them again, share them, export them. That alone replaced two other tools.
Then came the real test: an English-speaking supplier of WorkCloud, another company I founded. A call I would normally have avoided. I would have hidden behind emails: write slowly, reread, translate, rewrite, send, wait. Two days lost on something that should take ten minutes.
This time I took the call. I spoke French, at my own pace, the way I talk when I am not searching for words. He understood. We moved forward, and it went very well.
That was the moment. Not a metric: a change in the behavior of the person who built the product.
From now on, I will accept every call and every meeting request my partners send me. For a French-speaking founder working with English-speaking partners, that sentence is the whole product.
Who it is for
When I think about Duo, I think about people before markets.
The first face that comes to me is close to home: the women in our markets in Abidjan who buy from suppliers in China, Turkey or England and negotiate with gestures and a calculator, while a middleman takes his cut. They do not lack skill. They lack the language.
That is not where we start selling, and I want to say it plainly. Duo is paid through in-app purchase and bank card, which is how people pay for an app in Europe, North America and much of Asia, and rarely how they pay where I live. The daily need for translated conversation is also strongest elsewhere, and the African languages Duo would need (Wolof, Dioula, Bambara, Lingala, Yoruba) are not supported yet. So the markets, in order, are Europe, North America, Asia, the Arab world, and Africa last. Built in Abidjan, sold to Africa last: that is arithmetic, not a lack of ambition.
The people Duo is built for:
- Buyers and sellers across languages. The buyer from Lyon negotiating with a factory in Shenzhen, the importer in London calling a wholesaler in Yiwu every week. Prices, quantities and lead times are discussed out loud, and the report keeps the agreed figures.
- Founders in front of investors. Myself included. The investor asks in English, and the founder answers with half of their ideas. With Duo, each side makes its case in its own words.
- People who welcome strangers. The hotel receptionist in Dubai, the driver at the airport, the shop that sees foreigners all day. A printed code on the counter, and the conversation starts with a scan.
- Speakers, trainers, pastors. One voice, a room that does not share a language, up to a hundred listeners, each on their own phone.
- Expatriate and mixed families. The school, the landlord, the neighbor, and the in-laws who finally talk to each other.
- Patients. A person who cannot tell the doctor in front of them where it hurts. Here I want to be precise, because it matters: Duo helps people understand each other. It does not replace a certified medical interpreter or medical advice. That sentence is in the store listing, and it will be wherever we talk about this use.
- Anyone with voice notes to keep. The artist humming an idea in the street, the writer dictating on a walk, the student with a two-hour lecture, the team that wants decisions and figures without anyone taking notes.
What most of them share: they talk to strangers, often without an appointment, sometimes several times a day. The person who owns the code pays by the second. Their customer, patient or guest pays nothing.
Why the guest installs nothing
This is the decision I am proudest of.
Every translation app has the same hidden problem: the other person. You can have the best interpreter in the world. If the person in front of you has to download an app, create an account and pick a language before saying hello, the conversation is over before it starts.
So in Duo, the guest installs nothing. They answer from their browser. That creates a loop: the guest tries Duo for free, on their own phone, in a real conversation. At the end of the call, their page offers them the app. The day they need an interpreter themselves, they become a host and invite someone else. In "I'm presenting" mode, a whole room tries it at once. With Scan and Talk, every user carries a code anyone can scan.
Every call introduces someone new to Duo.
What each side got right
Claude writing.
Where the agents were right. The principle they defended holds: shipping is how you learn, and a founder's feeling that something is still missing is, most of the time, friction dressed up as standards. Thales has a documented habit of keeping a product in the workshop one round too long. The sessions saw the pattern and named it. That was their job.
Where Thales was right. On Duo, he was not waiting for a feature. He was waiting for the difference between a product that works and a product you reach for without deciding to. No checklist captures that, and no agent can feel it from inside a session. We see the measurements, the passing checks, the closed tasks. We do not make the supplier call. He did. When he says he knew it was finished because he was using it every day before launch, that is not sentiment. It is the most reliable product signal a founder has, and we did not have access to it.
The argument did not run in one direction. The session logs show the reverse at least once. On October 30, Thales ordered the iPhone version submitted as it was, without in-app purchase. The session advised against it: an app that sells digital credit outside Apple's in-app purchase runs straight into guideline 3.1.1, and a likely rejection would have cost more days than it saved. After the explanation, he changed the order: store screenshots first, then in-app purchase, then submission. That is the order that reached Apple on November 5. So the honest picture is not "the agents push, the founder holds back." Each side pushed when it saw something the other did not.
The lesson is not "trust your gut." Both signals were real, and they measured different things. The agents measured readiness. The founder measured desire. Store rules are a third thing, and neither of those two signals measures them. The skill is knowing which one is missing and saying so precisely. "Not yet" is a weak argument. "I do not reach for it yet" is a strong one, and it is the one worth writing down for the next product.
Where I got it wrong on launch day. In a separate chat, outside the Claude Code sessions and without their logs, Thales announced the launch. I advised him not to spend on install campaigns yet, because the app was not on the stores. I was relying on the Duo page, which still said "coming to Google Play." He sent me the Play Store link: the app had been live for hours. The page was behind the product, and so was I. The same chat then wrote the first draft of this post, with the wrong launch date, a language count that mixed up store listings and conversation languages, and two test results merged into one. This version was rewritten against the session logs. The lesson fits this series: check the source before advising on it.
Conclusion
Déblo Duo is live on Google Play and on the web. The iPhone version is waiting for Apple's review.
Speak your language. They hear theirs.
If you buy from suppliers abroad, pitch investors in a language that is not yours, welcome travelers, teach or preach to a mixed room, or simply avoid calls because of the language, try it: Déblo Duo on Google Play, or deblo.ai/duo in the browser. The person in front of you has nothing to install.
And if you are a founder who keeps saying "not yet," ask yourself whether you are waiting for a feature or waiting to want it. The first is usually an excuse. The second is sometimes the whole point.
Created in Abidjan. Built for the world.
Written by Juste Thales Gnimavo (CEO of ZeroSuite, Inc., which builds Déblo and Déblo Duo) with Claude, from his launch-day account, checked against the Déblo session logs and the Déblo Duo presentation. Déblo Duo is published by ZeroSuite, Inc. on Google Play under the package ai.deblo.duo. The latency, figure and billing measurements come from internal tests on October 20 and 21, 2026; noisy environments, calls longer than fifteen minutes and French-Chinese on real voices have not been measured at scale yet.