01Projet QSEIn production

do you think that with AI I could make the role-playing games I run with my students more immersive?

a teacher and trainer in Quality, Safety and Environment

The maker of educational role-plays.

Course material, or a few sentences of plain French, become a playable investigation: characters, documents, clues, marking rubric. Students question, unlock and write; the result comes back to the teachers.

From the first file to the first class
9 days
What's left of it today
Sold, delivered, runs without me
First file
8 April 2026

Enter the investigation Live product · designed, coded and shipped, solo

EM

Eleane Moreau

Responsable Qualité
EM

Eleane Moreau

Le lot L2025-F04-12 a été produit le 22/01 sur la Ligne 3. Pour la nature du verre, voyez la Maintenance.
Le labo a-t-il confirmé l'épaisseur du verre retrouvé dans la tartelette ?

« Quel lot est concerné, et sur quelle ligne a-t-il été produit ? »

The investigation window, playable right here: pick a witness, ask a question.

What the sentence meant

The role-plays already worked. The teacher was playing every character in person: “more immersive” meant witnesses who answer every student at the same time, with nobody standing behind them.

What was built

An investigation serious game. Students question characters played by AI, unlock clues, then write a report. The exercise is built around the concepts of the course: the academic material comes from the teacher, the product comes from me.

Then that scenario became one case among many. What was built next is the tool that makes these exercises: for any course, for any teacher in a school, with no training at all, because everything happens in plain language.

Three stages

ONE SCENARIO FOREIGN BODY four witnesses, three clues, written by hand A MACHINE THAT BUILDS read · analyse · method generate · check a document goes in, a playable scenario comes out A MACHINE YOU TALK TO “add a witness…” a course, an instruction, a file: a role-play THE SCHOOL in plain French, on every layer; a school hands it to its teachers 8 APRIL 2026 10 APRIL 26 APRIL TO 17 AUGUST
Dates from the repository: first file on 8 April, import pipeline and server on 10 April, Co-Pilote on 26 April, doctor and plans on 6 July, classes and notifications on 17 August 2026.

A course goes in, a complete exercise comes out

The teacher drops in whatever they have: course material, data, photographs. Five steps later, they open a playable scenario in the editor, with its characters, its clues, its documents and its rubric.

  1. 01Upload

    PDF, DOCX, images, CSV. Every file is read and sorted: course material, or a document to hand to the students.

  2. 02Analysis

    People, incidents, figures, timeline. Anything missing that would block the build is asked of the teacher before going any further.

  3. 03Method

    QQOQCCP, 5M (Ishikawa), DMAIC, Pareto: the teaching method chosen dictates what the scenario has to contain.

  4. 04Generation

    Characters, clues and root causes, documents: tables, charts, texts, images.

  5. 05Check

    A quality report with a score. Every flaw it finds is fixed in one click, with the Co-Pilote.

You talk to it, on every layer

Behind a scenario there are several hundred settings that nobody has to touch one by one. The teacher describes the change they want in plain French; a router picks the domain concerned, a sub-agent prepares the change, the teacher sees the preview, and applies, refines or rejects it.

THE TEACHER WRITES “add a maintenance technician and the clues they reveal” THE ROUTER picks the domain 7 SUB-AGENTS characters clues behaviours text documents metadata variables ONE CHANGE, TO APPROVE PREVIEW OF THE CHANGE Apply Refine Reject ten levels of undo
The editor's Co-Pilote, as docs/COPILOT.md describes it: the instruction quoted here is the example given in that documentation.

The school, the class, the feedback

A school opens an organisation account. Each teacher creates their classes, hands out an access code, and sees the marked reports come back. Students find their own in their space.

THE SCHOOL credit pot · licence the teacher creates their class they open THE CLASS access code 100 students at most one year the code THE STUDENTS PLAY they question, they write THE REPORT /100 marked against the rubric the report comes back marked the student keeps theirs thirty days email alerts
The organisation dashboard, the classes with their code, the waiting room, the reports and the email alerts all exist in the product; the figures are the ones in its code.

AI cost measured from the production logs between 10 April and 19 June 2026: a student session costs a median of $0.02, a full class between $1 and $3. Every class has a credit floor; a student who runs dry triggers an email to the teacher, who tops them up in one gesture.

What runs behind it

What the product does not do

The result

“More immersive?” became: a maker of educational role-plays that a school can hand to every one of its teachers. Sold, delivered, in production. The product runs on its own at the client's and is used in engineering schools, without me.