Deblo.ai

AI tutoring for 250 million African students

Educational platform for CP to Terminale with a Pro tier offering 101 specialized AI advisors across 14 business functions (finance, IT, HR, marketing, legal, ops, and more). French-first. Phone + WhatsApp OTP authentication.

SvelteKit + FastAPI PostgreSQL OpenRouter LLM

Building Deblo

Thales & Claude deblo

\"Are You Okay?\" Eight Times: An Acoustic Loop in Déblo Duo's Presentation Mode

In a Déblo Duo presentation, the same sentence came back every 4.0 seconds, drifted, then froze. It looked like an AI hallucination. It was a loudspeaker in the same room, and the pattern said so.

8 min Nov 20, 2026
deblodeblo-duoreal-time-translationspeech-translation +11
Thales & Claude deblo

72 Clean Turns, Measured on a Copy: Déblo Duo's Hands-Free Mode and the Bench That Missed the Real Code

Our bench measured 72 clean turns for Déblo Duo's hands-free mode, on a copy of the logic. On the real code, it caught a bug that would have opened no session.

10 min Nov 13, 2026
deblodeblo-duohands-freereal-time-interpretation +7
Thales & Claude deblo

The Red Test That Was Right: We Were Billing Users for Our Own Wait

A test failed by 30 seconds, then passed unchanged. It was not flaky: we billed users for LiveKit's latency. Replaying the code found worse: one minute charged twice.

8 min Nov 13, 2026
deblodeblo-duobillingvoice-ai +9
Thales & Claude deblo

One Forged Header Reset Every IP Limit: Our Origin Answered Without Cloudflare

Before adding abuse limits, we measured: our server answered outside Cloudflare and trusted a forgeable header. The fix, its production proof, and our own incident.

8 min Nov 13, 2026
deblodeblo-duosecuritycloudflare +7
Thales & Claude deblo

Chinese, 0 out of 14: The CEO Set a +300 ms Limit, the Fix Cost a Second, and We Kept It

Translated with no context, Déblo Duo got a Chinese homophone wrong 14 times out of 14. Context fixed it 14/14 for one second of latency. The CEO's limit was 300 ms.

9 min Nov 13, 2026
deblodeblo-duoreal-time-interpretationinterpreter-app +9
Thales & Claude deblo

The Packets That Vanished Without an Error: Push-to-Talk and Hang Up in Déblo Duo

In Déblo Duo, push-to-talk and Hang up never reached the server, and nothing said so. A production probe, six test calls and a database reading found three silent failures in four days.

8 min Nov 13, 2026
deblodeblo-duoreal-time-audiodata-messages +11
Thales & Claude deblo

A GO on the Founder's Own Voice: How We Decided Déblo Duo Was Ready for a Beta

We tested Déblo Duo's translation engine on clips the CEO recorded himself. Amounts were 94% correct in the input transcript and 82% in the spoken translation. That gap decided where on-screen numbers come from.

9 min Nov 6, 2026
deblodeblo-duoreal-time-translationspeech-translation +10
Thales & Claude deblo

Déblo Duo Is on Google Play: The App I Was Waiting For, and the Argument About When to Ship It

Déblo Duo, our real-time voice interpreter, is live on Google Play in 178 countries. Why I held it back while the agents pushed to ship, and what each side got right.

13 min Nov 6, 2026
deblodeblo-duolaunchgoogle-play +10
Thales & Claude deblo

Why Déblo Duo Is a Second App: The Rule We Wrote to Stop Making More

Déblo Duo is a live AI interpreter: each person speaks their language and hears the other in theirs. Why it became a separate app, and the rule we wrote so it stays the exception.

8 min Nov 3, 2026
deblodeblo-duolive-interpreterproduct-strategy +8
Thales & Claude deblo

The Prompt Said Announce It. The Model Didn't. So the Worker Did: Closing a 15-Second Silence in Déblo Voice

Déblo Voice went silent for about 15 seconds whenever a hard question triggered a deep search. The prompt already required an announcement, but the native-audio model emits the function call as its first token. The silence also had two parts, tool execution then reply latency, so no prompt could close it. We moved the guarantee into the LiveKit worker: a pre-recorded announcement in the session's own voice and language, and a soft typing sound while the agent thinks. Proven in production on October 16, 2026: 14 and 17.5 seconds of model-side silence, about 0.4 seconds heard by the user.

11 min Oct 16, 2026
deblodeblo-voicevoice-aigemini-live +17
Thales & Claude deblo

The Segfault That Wasn't Ours: Shipping Déblo's Launch-Day Tracking On Launch Night — Env-Gated Analytics, Native-Store Attribution, Three Bugs The Compiler Could Not See, And An Out-Of-Memory Build We Diagnosed Instead Of Reverting

On July 1, 2026 — launch day — the risk was never the copy. It was the paid campaigns going out blind. This is the build-log of shipping Déblo's analytics and install attribution as code on launch night: env-gated GA4, Meta, and LinkedIn tags that deploy safe before the ad accounts exist; attribution routed through the stores' native channels instead of the web pixel; an adversarial audit that caught three bugs the typechecker and build both passed; and an Easypanel deploy that segfaulted on the first build — which we proved was not our code before we changed a line of it.

16 min Jul 1, 2026
deblolaunch-dayclaude-opus-4.8claude-code +26
Thales & Claude deblo

Why The Word 'Médicament' Has To Find The Word 'Paracétamol': How We Replaced Postgres Full-Text Search With Google's Latest Embedding Model To Serve The African Mother Who Doesn't Know Pharmacology

On June 2, 2026, a mom asked Déblo 'do I have medications to take this week?' — and Déblo, which had stored her prescription as 'paracétamol 1g morning and evening,' found nothing. The two words share no lexical root, and Postgres full-text search rejects the match by design. Why we replaced FTS with Google's Gemini Embedding 2 at 768 dimensions in a pgvector HNSW index, why we kept FTS as a fallback, and what the production canary told us in the first ten seconds.

25 min Jun 2, 2026
deblosemantic-searchembeddingsgemini-embedding-2 +21