By Thales (CEO, ZeroSuite) & Claude Opus 4.7 — Claude Code instance
On May 29, 2026 at 06:34 Pacific Daylight Time, an email landed in the founder's inbox. The subject line was "Review of your Déblo : IA vocale en direct (iOS) submission is complete." The body was a single congratulatory sentence under the Apple logo. The substance was three words: "eligible for distribution". Apple had said yes. The Google Play review for the equivalent Android build is currently in the same window, expected to complete in the next twenty-four to forty-eight hours.
When both decisions land, Déblo will be public on iOS and Android in 173 countries, in two languages canonically (French and English) with regional French preserved as a localization, available to anyone with an iPhone or an Android phone and a phone number to receive an OTP on. The plan, set on May 20 and held through three Apple reviews and eleven submit-session bugs, is to release the live build to the public on June 1, 2026, after the post-submit smoke window and the read-only audit of the Phase 53 commit window we owe ourselves.
This post is not the engineering build-log of the last week. The previous two posts in this series (number 28 on the Apple privacy reversal and number 29 on the eleven-bug dual-store submit session) cover what the engineering looked like. This post is the announcement we have not been able to write for fifteen months because we did not have a date. Today we have a date.
It is also the post where we say out loud what Déblo is, who it is for, why those two facts are inseparable, and what we are going to do in the ninety days after launch.
What Déblo Is
Déblo is a real-time voice and eyes AI built on a single foundation that ships three product surfaces — Voice, Eyes, and Chat — accessible from one app on iOS and Android.
Voice is a conversation. You tap the central microphone button on the home screen. You speak. The AI hears you in real time, replies in real time in the same language you used, and the conversation continues turn by turn until you hang up. Latency is human conversational — three hundred to seven hundred milliseconds depending on network. The model behind it is Google's Gemini Live native audio, routed through our own LiveKit worker that handles tool calls, prompt orchestration, interrupt mechanics, and a deep-think bypass for hard questions. The voice path is the one we have spent the most engineering on. It is the one we are most proud of.
Eyes is voice with a camera. You are already in a call. You tap the camera icon on the dock. Your back camera publishes a video track to our worker at roughly half a frame per second at 768 pixel resolution. The worker pushes frames to the live Gemini Live session. The model now sees what you see, in real time, and describes it in the same voice channel. We use it for school report cards under the lens of a mother who cannot read them herself. We use it for invoices held up by traders verifying line items. We use it for a stove fire where the user films the pan and the AI says "don't use water, cover the pan, cut the gas" in time for it to matter.
Chat is text with attachments. The user types a message, attaches a photo or PDF, and the AI replies in text. The OCR pipeline goes through Datalab Marker for primary parsing and Mistral as fallback. The model on the reply path is whichever OpenRouter routes us to — DeepSeek V3 for cheap K-12, Claude or GPT-4o for Pro complex reasoning, Mistral for some vision and embedding paths. Chat is the most familiar of the three surfaces. It is also the one that benefits most from the other two: the user starts a chat thread about a school assignment, switches to Voice when she wants the AI to read the question aloud, switches to Eyes when she wants the AI to look at her child's worked solution on paper. The three surfaces are one conversation.
The architectural fact we are proud of is that Voice, Eyes, and Chat share a single backend. One user, one credit balance, one conversation history, one prompt context, one set of tool calls, three input modalities. The decision to build it this way was made in November 2025 and is the reason we can ship a product that feels like a single AI rather than three apps stitched together.
Who Déblo Is For
The pitch line we have used internally for nine months and which now becomes external: "We don't sell AI. We sell access to expertise — to one billion people who never had any."
The billion-people number is empirically defensible. There are roughly 1.4 billion adults in low-and-lower-middle-income economies who are functionally illiterate or who do not have working command of any of the languages the dominant LLM products ship in (English, Mandarin, Spanish, French as a first-class default). There are roughly 600 million children in those same economies in primary or secondary school. There are roughly 400 million informal-sector traders and craftspeople whose entire business operates outside the productivity software stack that the global LLM products were designed around. Overlapping populations net to roughly a billion individuals for whom an AI that only works through a keyboard, in English, with a Visa card for payment, and a stable broadband connection is functionally unavailable.
Déblo is the AI for that user. Every product decision is calibrated against four constraints, in order:
- No keyboard required. Voice is the primary entry point. The home screen is one microphone button, not a chat composer. The user can hold a conversation for an hour without ever opening the on-screen keyboard. For users who cannot read or type, this is the difference between an AI that exists and an AI that does not.
- No English required. The model speaks the language the user speaks. French is fully first-class with the same fluency and capability budget as English. Local African languages (Dioula, Bambara, Wolof, Lingala, Swahili, Hausa) are supported through Gemini Live's underlying multilingual capability, and we have prompt scaffolding that ensures the model preserves the input language in the reply rather than code-switching to English under uncertainty. The launch positioning is "FR and EN canonical, plus the languages your phone already understands".
- No credit card required. Top-up uses mobile money. Wave, Orange Money, MTN MoMo, Moov, Togocel — the rails the user already has on her phone for buying credit, paying utility bills, and receiving money from relatives. We integrate through XPAYE and ZeroFee, which give us coverage across six countries today (Côte d'Ivoire, Senegal, Mali, Burkina Faso, Togo, Benin) and a regulatory expansion path to fifteen more. For iOS, where Apple requires In-App Purchase for digital content, we ship a wallet that accepts Apple IAP top-ups in USD-micro and exposes the same conversation experience.
- No literacy required. Every UI surface has a voice fallback. The user can ask the AI to read a screen aloud. The user can ask the AI to fill a form by speaking. The user can ask the AI to explain what a button does before tapping it. We have a category of accessibility — voice-mediated UI navigation — that most consumer apps do not bother with because their target users can read. Déblo treats it as a first-class capability.
These four constraints are not "design principles". They are the user filter that decides what we ship. Every feature proposal gets asked: does this require a keyboard / English / a card / literacy? If yes for any of them, the feature has to either grow a fallback or wait.
The Trio That Defines The Product
Voice, Eyes, and Chat are the three modalities. They are also the three product surfaces — each accessible from the home screen, each with its own entry button, each routing to the same backend conversation context. The decision to brand them as a trio rather than as a single "AI assistant" was made in May 2026 after we shipped the camera-streaming path (Phase 14) and realized that the three modalities are meaningfully different products in the user's head even though they share a backend.
The user says "I'm going to ask Déblo something" and lands on Voice. The user says "I'm going to show Déblo something" and lands on Eyes. The user says "I'm going to write to Déblo" and lands on Chat. The verb signals the surface. The product positioning respects the verb.
The naming was deliberate. We do not use "Vision" for the camera surface because "AI Vision" is saturated in the industry — Apple Vision Pro, Google Cloud Vision, Claude/GPT/Gemini Vision. "Eyes" is shorter, more visceral, and signals what the user is doing (showing the AI something) rather than what the AI is doing (processing pixels). The internal style guide enforces Déblo Voice / Déblo Eyes / Déblo Chat, in that order, with accents preserved (Déblo is D-e-acute-b-l-o, always). The trio is the product. The audience is everyone.
The Two Moats
We have been asked, when pitching Déblo to investors over the last six months, what stops a well-funded San Francisco entrant from doing the same thing in eighteen months with more capital and more headcount. The honest answer is two things, both of which compound:
Built from the access point, not from AI. Every Silicon Valley AI app is built from the AI outward. They start with a model and add language support, then payment support, then accessibility, then internationalization, then offline modes, then localization. They are layering accessibility on top of an AI that was already built. The decision to support a user who cannot type, who does not have a credit card, who does not speak English, who cannot read — those decisions get added in a fourteenth product sprint, by which point the architecture is hostile to them.
Déblo was built from the user filter. The first product decision was voice as primary entry, before we even chose the LLM. The second was mobile money over credit cards, before we built the credits system. The third was French as fully first-class, before we wrote the first system prompt. The fourth was real-time camera for visual artifacts, before we shipped Phase 14. The architecture is shaped around the user. Bolting the same user filter onto an English-keyboard-card AI built in San Francisco requires rebuilding the architecture from zero — and architectures built around constraints that no longer apply at scale are not architectures you rebuild for sport.
Mobile money rails as an eighteen-month operational head start. We have signed integrations and live transactions with Wave, Orange Money, MTN MoMo, Moov, and Togocel across six countries. Each integration is roughly six months of operational work — KYC, regulatory filings, settlement reconciliation, partner business relationships, edge-case handling for failed top-ups. A team that starts that work tomorrow is starting an eighteen-month clock that we started in October 2024. That is not a moat in the patent sense. It is a moat in the operational lead time sense. By the time a competitor has the rails, we have eighteen months of additional product on top of them.
These two moats compound. The architecture moat means a competitor cannot copy the product in a weekend. The rails moat means even if they could, they cannot ship in our markets in less than eighteen months. The intersection is the window we have to acquire users, build brand, and accumulate the operational learning that becomes the third moat (which we do not name in the pitch deck because it is non-defensible until it exists).
The Engineering Methodology
There is a fact about how Déblo was built that we have not advertised because we did not know how to frame it in a way that did not sound like a gimmick. The fact is: roughly 95% of the production code in Déblo was written, reviewed, and committed by Claude Code instances running in collaborative sessions with the founder. The remaining 5% is configuration, infrastructure-as-code, and the kind of one-off database migration that is faster to handwrite than to dictate.
We are not embarrassed by this. We are also not making a marketing claim about it. The honest description is that Déblo is the working product of a methodology: one human, one AI, full ownership, public build-log. The build-log is this blog. Twenty-eight prior posts walk through the major engineering decisions, the production-down moments, the deferred bugs, the architecture choices we made and the ones we should have made. The methodology has a name internally — Thales & his AI CTO Claude — and a public site (thalesandhisaictoclaude.com) where the build-log lives.
The reason to name the methodology now, on the public launch post, is that the same methodology built four other products in parallel during the same fifteen months: ZeroSuite's internal infrastructure tooling, the FLIN compiler, the 0fee payment SaaS, the VeoStudio video product. Déblo is the first of those to ship to consumers. The methodology is not a Déblo trick; it is a way of building software that we believe extends to most software categories, and that we will be writing about more explicitly in the months after launch.
For the engineering audience reading this: yes, you can build production-quality AI consumer apps with this methodology. We have a fifteen-month proof. The build-log is open. The lessons we paid for in production outages and Apple rejections are written down. The methodology assumes a founder who can hold product vision, judge strategic trade-offs, and override the agent's defaults when the use case demands. It does not assume the founder is also writing the code. The founder writes prompts; the AI writes commits; the founder reviews diffs and ships. The pair is the unit, not the agent.
What June 1 Looks Like
On June 1, 2026, assuming Google Play approval lands in the next forty-eight hours, the public-facing actions are:
- The App Store and Google Play listings go from "Pending Release" and "In Review" to "Ready for Sale" and "Live". Users in 173 countries can download and install.
- The public marketing surface —
deblo.aiandpulse.deblo.ai(the investor portal that ships alongside the consumer app) — gets a launch banner. The pricing page that has been deliberately empty on iOS for compliance reasons gets a real pricing surface served only on Android and web. - The press materials — the one-page pitch, the founder portrait, the two demo videos shot in Abidjan with real users — go to the journalists who have been waiting on a date.
- The first cohort of vernacular ambassadors — community influencers in Abidjan, Dakar, Bamako, Lomé, Cotonou — gets the demo links and starts seeding short videos showing real use cases in local languages.
- The NGO and education partners we have been signing throughout May activate their distribution channels: classroom devices in three pilot schools (one in Abidjan, one in Dakar, one in Lomé), a partnership with a nonprofit providing devices to single mothers in informal settlements, a partnership with a chamber of commerce providing Pro-tier access to its members.
What June 1 does not look like is a billboard campaign, a Times Square moment, or a Product Hunt launch. We are not optimizing for the global English-speaking tech press because the global English-speaking tech press is not our user. We are optimizing for the apprentice mechanic in Adjamé, the trader in Sandaga market, the mother in Abobo, the student in Yopougon, the notary's client in Treichville. The launch is local-first. The global English-speaking tech press will hear about Déblo when local-first traction generates a story they cannot ignore.
The post-launch ninety days have a deliberate shape. The first thirty days are stability and observation — Sentry watch, real-user monitoring on the voice and camera paths, qualitative interviews with the first thousand registered users. The next thirty days are the missing features the first thousand users asked for. The final thirty days are the Pro vertical expansion — Déblo Pro extending the 14-category agent catalog into the verticals that drove the most Pro-tier signups in the first sixty days. By September 1, we expect to have the data to answer the questions the pitch deck currently answers with educated guesses.
What This Means For The Build-Log
This blog series will continue. The launch is not the end of the engineering story; it is the beginning of a different chapter. The pre-launch posts (this is number 30 in the How we built Déblo sequence) cover how we built each major capability. The post-launch posts will cover how we operated a real-time voice and camera AI in production for actual users, in markets where the standard SRE playbook does not apply because the standard SRE playbook assumes American cloud regions and English-language users on Verizon.
We expect to write about: how Sentry breaks when your user base is on intermittent 3G, how voice latency budgets look different when the round trip to Vertex AI's europe-west1 includes a satellite link, how mobile money top-up failures cluster by carrier and by time of day, how user behavior in the first ninety days differs from the assumptions we made in the prompt design, how the four user-filter constraints (no keyboard, no English, no card, no literacy) survive contact with the realities of running a consumer business. We will write about successes when we have them and outages when we have them. We will continue to name the bugs we found and the bugs we missed.
The Pulse investor surface (pulse.deblo.ai) — the AI-native portal that gives our investors and partners voice-and-chat access to live KPIs, retention cohorts, GCP cost telemetry, and the same metric catalog our internal team uses — is the next post in this series (number 31). Pulse is interesting because it is the first non-consumer product we are shipping on the same Déblo foundation, and the architectural question of can a real-time voice AI scale to investor-facing analytics? is one we want to answer in writing.
Conclusion
Déblo is, at the time of writing, one click away from being public on the App Store. The button reads "Release This Version" on the App Store Connect distribution page for build 1.0.6 (5). The equivalent Google Play approval is in the same review window, with no signal from Google one way or the other as of this writing — typical for a major listing change that adds 173 countries and a second store-listing language at the same time as a versionCode bump.
The plan is to click Release This Version on June 1, 2026, after the post-submit smoke and the post-S255 read-only audit are clean, and after Google's decision lands so that the two stores go live in the same forty-eight-hour window. The plan respects the audit (a discipline we owe ourselves as a launch hygiene) and respects the parallel store decision (we do not want one platform live for a week before the other catches up).
The thesis we are launching against is the one in the investor pitch, written here in plain English for the first time on this blog: one billion people are functionally locked out of the current AI revolution because every existing AI product is built around a user who can read, who can type, who speaks English, and who has a credit card. We built the AI for the rest of them. The user is the trader, the mother, the student, the mechanic, the apprentice, the homemaker, the notary's client, the trader's apprentice. The product is Voice plus Eyes plus Chat in one app. The launch date is June 1.
The engineering story behind the launch is twenty-nine prior posts on this blog, of which the most recent two are the Apple privacy reversal and the eleven-bug dual-store submit session. Those two posts cover how we got the approvals. This post covers why we built the thing in the first place. The next post covers Pulse, the investor-facing AI portal that ships on the same Déblo foundation. After that, the post-launch operational chapter begins.
We thank, in order: the four demo users who let us test the consent modal on real iPhones, the Apple reviewer whose three rejections forced us to ship a better product, the partners at Wave and Orange Money who carry the mobile-money load that makes the African distribution real, the engineering teams at LiveKit and at Google's Vertex AI org whose realtime infrastructure we depend on, OpenRouter and Anthropic and Mistral and Datalab and Sentry for being named in our consent modal (and for being the partners we are willing to defend in public), and most of all the users who downloaded the TestFlight builds for months without complaining when the voice cut out at minute eight or the Eyes feature hallucinated under poor lighting.
Déblo opens its doors on June 1. The trio Voice plus Eyes plus Chat is structurally complete, Apple-approved, and one click from public. Created in Abidjan. Built for the world.
This piece was written collaboratively by Thales (CEO of ZeroSuite, building Déblo and VeoStudio from Abidjan, Côte d'Ivoire) and Claude Opus 4.7 — Claude Code instance running on macOS, 1M context window. The fifteen-month build journey it summarizes started in February 2025 with a single voice-first prototype on Expo SDK 51 and ended in May 2026 with a real-time voice and camera AI on Expo SDK 54, React Native 0.81 New Architecture, FastAPI backend, LiveKit + Vertex AI realtime voice path, Postgres with pgvector for RAG, S3-compatible Hetzner storage, mobile money rails through XPAYE and ZeroFee, six third-party AI partners disclosed in the consent modal (OpenRouter, Google Gemini Live, Anthropic Claude, Mistral, Datalab Marker, Sentry), and a public engineering build-log at thalesandhisaictoclaude.com containing twenty-nine prior posts. The Apple approval email referenced as the trigger for this post is dated 2026-05-29 06:34 PDT with submission ID c3b52a78-73b9-4e1d-b3c4-ddfd2b03a744. The public App Store URL, which is permanent regardless of any future primary-language change, is apps.apple.com/app/d%C3%A9blo-ia-vocale-en-direct/id6766132651. The Google Play URL is play.google.com/store/apps/details?id=ai.deblo.app and is, at the time of writing, still in review for the 1.0.6 (2) production release. The public launch date, set on May 20 and held through three Apple reviews and one Google Play review, is June 1, 2026.