QDNAAdvisory and architecture for LLM inference and training platforms, on-premises or hybrid

apprendre-ia.qdna.fr: learning AI at work, from beginner to expert

A living case study: a free learning site that teaches how to work with AI, starting with the tool most employees already have, Microsoft 365 Copilot Chat. Designed, deployed and operated by QDNA, launched in September 2026.

QDNA case study · Published on · Updated on

Short answer. apprendre-ia.qdna.fr offers a free course of 24 lessons in four levels, Beginner, Intermediate, Advanced and Expert, on using Microsoft 365 Copilot Chat at the office. Each lesson takes fifteen to twenty minutes and ends with a corrected quiz and a proof of mastery, the learner's final prompt and the result obtained. A prompt composer guides the writing field by field, and a signed Open Badge attests each level. QDNA, an architect of sovereign AI platforms, on-premises, hosted or hybrid, designed it, deployed it and operates it: the same chain as gironde.qdna.fr, applied this time to training.

The site in figures

CourseMicrosoft 365 Copilot Chat, 24 lessons in four levels of six lessons, in French and in English
Companions10 prompt sheets by job, a glossary of 51 terms, a prompt composer, one page per level and one page per badge
Lesson formatobjective, cold attempt, concept with a diagram, worked example, exercise with the composer, pitfall, three-question quiz, proof of mastery, FAQ
Accountoptional: the whole site can be read without one; sign-in is a link received by email, with no password, and stores only the address and the progress
Statusin production since September 2026, published by QDNA, served at the network edge (Cloudflare), free of charge

What the site does

The starting point is simple: anyone with Microsoft 365 at the office already has Copilot Chat, with no additional licence. The site teaches how to get useful work out of it, without jargon: write a complete prompt, provide the right source, check an answer, then iterate, chain prompts, analyse a file, equip a team and, at the last level, write a usage policy and train colleagues. The four levels follow the UNESCO AI competency frameworks and the proficiency levels of the European DigComp framework, and each lesson targets a named skill: framing, judging or steering.

Each lesson opens with a direct answer under the title, then a concept illustrated by a diagram, a worked example that puts a weak prompt and a strong prompt side by side, and a pitfall to know: hallucination, confidential data, doubtful source, overconfidence. The quiz corrects immediately, and the lessons of the next level reuse a notion from the previous one, which is the simplest form of spaced repetition.

The pedagogy: proofs of mastery, badges, composer

The proof of mastery replaces the grade. At the end of each lesson, the learner pastes the prompt actually written in Copilot Chat and the result obtained; this proof is kept in the learner's portfolio and counted towards the level badge. The site shows no leaderboard, no daily streak and no global score: progress is read by skill and by proofs submitted.

The badge of a level is issued once every proof of the level is submitted, the scheduled reviews are passed and the synthesis test is completed. It is a credential in the 1EdTech Open Badges 3.0 format, signed by QDNA, which carries the date, the level and the criteria, but neither the learner's name nor email address. Anyone who receives it can verify it on the site from its identifier. The site states that it is neither a diploma nor a certification: it attests progress, nothing more.

The prompt composer assembles a prompt from the ingredients taught, objective, context, expectations, source, constraints and verification, with optional boosters such as "cite your sources and flag uncertainty" or "offer two variants". It is embedded in the lessons where it belongs, with the fields and criteria specific to each, and nothing typed leaves the browser.

The architecture: Cloudflare Pages, Functions, D1, Email Service

The foundation is the one used by qdna.fr: static HTML pages generated by Python scripts from Markdown files, one per lesson and per language, with JSON-LD structured data (course, learning resource, quiz, FAQ, breadcrumb), canonical, hreflang and visible dates equal to the declared ones. A validation gate refuses any incomplete publication, and the generated output is never edited by hand.

The learner account fits in a few Pages Functions: sign-in receives an address, creates a single-use token valid for fifteen minutes and sends the link through Cloudflare Email Service; the response is the same whether the address exists or not. The session is a signed cookie, and a D1 database holds five tables: users, sign-in tokens, progress, proofs and badges. The badge is a verifiable credential signed on the Worker side, and a single deletion request erases everything at once, in line with the GDPR. No password is stored, and no data beyond the address and the progress.

Why this is a QDNA case study

Training users is part of putting an AI platform into production, in the same way as choosing the hardware or the model: a tool nobody knows how to frame, check or steer does not deliver the expected service. This site shows how QDNA approaches the question, with a pedagogy built on practice and proof rather than reading, and the same demand for accuracy as on its other sites, every page citing its sources. The site is published by QDNA and its structured data references the qdna.fr/#org organisation as publisher.

Visit the site: apprendre-ia.qdna.fr, the Copilot Chat course, the four levels and the badges.

Sources

Let's discuss your project

A no-commitment call to frame your platform, team training included.

Book a call