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.
Building 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.
Pulse: How We Replaced The Pitch Deck With A Real-Time Voice AI That Investors Can Ask Direct Questions To — On The Same Foundation As The Consumer Product
Pulse is the investor-facing surface of Déblo, built on the same FastAPI backend, same LiveKit worker, same Gemini Live model. Magic-link HMAC RBAC, thirty-five voice tools plus three helpers, a Postgres materialized view for retention math, the radical-minimalism home redesign, and the one-shot action tools prompt rule. Due diligence as demo.
Déblo Opens Its Doors: After Fifteen Months Of Building And Three Apple Reviews, The Real-Time Voice And Eyes AI We Made For The Billion People Without Access To Expertise Is About To Be Public
On May 29, 2026, Apple approved Déblo for distribution. The launch post that names the thesis — one billion people locked out of AI by keyboards, English, credit cards, and literacy — the two moats, the trio Voice plus Eyes plus Chat, the engineering methodology, and what June 1 actually looks like.
Naming The Six Partners: How An Apple Rejection Forced Us To Reverse The Decision To Hide Our Stack, And Why The Reversal Was The Right Product Call
Apple's three-guideline rejection on build 1.0.5 forced us to reverse the session-178 CEO decision to hide the AI stack. Why we now name OpenRouter, Google Gemini Live, Anthropic Claude, Mistral, Datalab Marker, and Sentry in the consent modal before the Accept button — and what the reversal taught us about disclosure surfaces.
Eleven Bugs Between Submit And Ship: A Five-Hour Dual-Store Submission Session, Walked Through Bug By Bug, From RCT-Folly Podspecs To Sixteen-Kilobyte Page Sizes
Eleven distinct bugs found and shipped in a single five-hour dual-store submit session, from RCT-Folly podspecs under Expo SDK 54 to a sixteen-kilobyte Android page-size warning. Bug by bug, what broke, what the fix looked like, which three required persistence-layer follow-ups, and which one we deferred cleanly to versionCode 3.
Step Zero Wasn’t Enough: How Validating A Constructor But Not The Runtime Took Down Every Déblo Voice Session The Hour We Shipped Real-Time Camera Streaming
Phase 14 shipped Déblo Eyes — real-time camera streaming over LiveKit to Gemini Live native audio. The first deploy took down every voice session in production within ninety seconds because our Step 0 had validated the constructor without exercising the runtime path. The build log of how Déblo got eyes, what an incomplete pre-flight check cost us, and which polish items we shipped versus deferred.
The Em-Dash That Killed Production: How One Marketing Tagline In An HTTP Header Took Down Déblo’s Chat For 24 Hours
Two days before App Store submission, Déblo’s entire chat product silently broke. No spinner, no toast, no error in the UI — just dead air. The 24-hour outage came down to a single « é » in an HTTP header value raising UnicodeEncodeError before any request to OpenRouter ever left the backend. The post-mortem of a false hypothesis, a Sentry trace, and a 6-line fix that unblocked the launch.
Six Hours From Empty Page to Apple Review — How We Submitted Déblo to the App Store, Live
Live walkthrough of submitting Déblo to the iOS App Store in six hours: what Apple’s validators rejected (a Unicode superscript), what we corrected (a Promotional Text wasted on third-party brands), and the iOS ASO mechanics almost everyone gets wrong.
Trust the Model, Tell It Less — How We Compressed Déblo’s System Prompts by 38%
Eight hours of prompt compression at the CEO’s directive: five system prompts shrunk from 138K to 85K characters (−38%), 15 verbatim French templates deleted, pricing context plumbed per country, and Déblo’s identity opened beyond Africa to AP/SAT, GCSE/A-Level, and IB.
The Day Déblo Refused a Correct Answer — Twice
A production trace showed Déblo K12 rejecting a Terminale student’s correct answer twice in a row. Eight hours of forensic work, four commits, one A/B-tested model rotation, and a 6-model benchmark later, the math tutor was fixed. Here is exactly what broke, what we changed, and what GPT-5.4-mini’s surprising failure on the Socratic test taught us about picking models for educational AI.
Web Claude Found the Bug. Then It Almost Made It Worse.
How a 270-line voice prompt for Deblo's Ultravox tutor was producing the same scripted greeting every call. Web Claude diagnosed it perfectly, then prescribed a fix that would have doubled the prompt size with non-existent backend hooks. The filter that kept the diagnosis and rejected the prescription.
Why I Had to Correct Web Claude Twice on Deblo's Home Page Strategy
How a 48-hour conversation with Web Claude almost led Deblo into the "ChatGPT for Africa" generalist trap, and why founder market knowledge had to override AI strategic suggestions twice. Plus: what serious AI-augmented product teams actually look like.