"Just another ChatGPT wrapper"? What is really inside KAÏDO

"Just another ChatGPT wrapper". I have read it several times under our Instagram posts. So yes, it is true: our prototype uses GPT-Live, OpenAI's voice model.

But KAÏDO is not just an AI plugged into an object. Let me explain all the work we have been putting in for months.

• KAÏDO never gives the child the answer. A language model, on the other hand, is built to answer whatever question it is asked. Extra difficulty: the voice comes straight out of the model, there is no text to check before the child hears it. So we built 3 layers: a written, versioned set of core rules, AI-free safeguards that read every spoken word in real time, and a second model that reviews each exchange and steers what comes next. Extra extra difficulty: a child can be hard to understand. We put in place what it takes to listen even when articulation is poor.
• A teaching approach written and refined with teachers and speech therapists, not by the model: more than 470 lesson sheets from first to fifth grade, built from the official French curriculum. Each sheet says how to guide without giving the answer: progressive hints, common mistakes, prerequisites. On top of that: spaced revision, tracking the topics they still struggle with, and going over the day's lesson.
• Each child has their own teaching approach. We decided how to help the child, not the model: when a child is stuck, we choose among several teaching moves (a small hint, breaking it into steps, encouragement, an example on another case, an analogy, a question...) based on what we know about them: their mastery of each notion, their recurring mistakes, the prerequisites they are missing. Two children stuck on the same exercise will not be helped the same way.
• What the model might do badly, we do not entrust to it: the first meeting with the child is written and scripted, not improvised. KAÏDO also knows how to stay quiet: it lets the child work and only steps in when called.
• We never keep the child's voice, nor their conversations. We only remember where they stand (their strengths, what resists, their progress) to adapt how we help. The parent gets an update every week, to be reassured, not to monitor.
• We do not assume, we check: hundreds of sessions tested, from first to fifth grade. Every slip we observe becomes a rule, then a test replayed with every version.
• A playful physical object, screen-free, designed in France. Electronics, firmware and iOS and Android parent apps developed in-house.

GPT-Live understands the child and answers. KAÏDO helps them understand, decides what can be said and what must be left unsaid. And if tomorrow a European model does better than GPT-Live (Mistral AI, shall we talk?), we will switch. The promise stays the same: guide the child, never answer for them.