T-TWICE : Mathematical Reasoning Engine

Team name: T-TWICE

Use of AI tools : We used Claude (Anthropic) for structuring arguments, drafting, and stress-testing our reasoning. The POC runs on Scaleway Paris infrastructure (sovereign, GDPR-compliant), with a language model migrated from DeepSeek to Scaleway's hosted models for data sovereignty.

External feedback & contributions :

• Thierry Jeantheau – Researcher & lecturer, Université Gustave Eiffel – Mentoring, problem validation (L1 mathematics failure rates), discussion on GDPR compliance and algorithmic bias

• M. Seuwouo – Mathematics teacher, secondary school – Real classroom testing across multiple tutoring groups

• Immanuel Shadrach – Participant – Raised the question of motivating reluctant students → shaped the engagement system; flagged AI errors in complex mathematics → reinforced the requirement for human oversight

• Amadou Gaye – Participant – Challenged whether a tool that never gives answers runs counter to real usage patterns → shaped the progressive engagement system design

• Nicol D. – Participant – Pushed for a minimum viable version → guided POC feature prioritisation

• Parents of students (via Candis) – Field users – Expressed an unexpected need for autonomous parental follow-up → extended the inclusive dimension beyond the institutional frame

Initial contribution: Please provide the link (URL) to your initial contribution on the platform: (Opens in new tab)T-TWICE : Mathematical Cognitive Reasoning Engine

Final contribution: Our initial contribution already carried the full ambition of T-Twice: an AI tutoring tool that guides mathematical reasoning without replacing it. What Phase 2 revealed was not a flaw in the vision, but a gap in the experience. The tool worked, but it was not yet designed to make every type of student want to stay.

The feedback from Amadou Gaye and Immanuel Shadrach crystallised this: a student who is not intrinsically motivated will not persist with a tool that refuses to give answers, however pedagogically sound that refusal is. Nicol D. pushed us further — what is the minimum viable version that actually delivers? These questions forced us to go deeper into the product itself.

The result is a complete engagement system built around three pillars (student, professor, and interface) designed so that the student builds real reasoning without realising they are applying the cognitive mechanisms that make learning last. The goal is not compliance. It is genuine pleasure in practising mathematics. This system is documented and testable: https://silly-sprite-52ceaf.netlify.app(External link)

The inclusive dimension goes further than we initially proposed. With the consent of all parties, T-Twice's neuroadaptive system learns how each student (with or without a diagnosed condition, dys profiles or otherwise) actually behaves, and adapts in real time so that even this student feels at ease and motivated to continue. The full architecture is here: https://comforting-longma-1fcb29.netlify.app(External link)

On GDPR and AI Act compliance, Kankeu and the team worked through concrete bias questions with Thierry Jeantheau. The model migrated from DeepSeek to Scaleway Paris for full data sovereignty. The legal and ethical framework is detailed here: https://storied-centaur-208d7c.netlify.app(External link)

The POC has been tested informally with a small group of students from Candis and Johanna's tutoring sessions, whose anonymised interactions are accessible directly in the platform. We do not claim statistically significant results at this stage — the sample is too small to draw conclusions. What is implemented and testable in the POC versus what is planned, and what we observed, qualitatively, are documented here: https://taupe-sable-d02453.netlify.app(External link) . Large-scale empirical validation is the first milestone our roadmap targets once the project is funded, because running rigorous trials at scale has a real cost, and we want to do it properly.

Finally, to connect the prototype with a realistic path forward, a deployment, inclusion and sustainability note is available here: https://celadon-brigadeiros-5937cf.netlify.app(External link)

It explains why the need is real, how T-Twice differs from existing AI tutoring tools, who it is designed for, and how the project could scale responsibly in Europe and in African educational contexts.

Demonstration URL : https://ttwiceda133040-frontend.functions.fnc.fr-par.scw.cloud/(External link)

Login credentials:

User: jury

Password: zfi6pONdIfAh5avU

The POC is currently live and can be tested directly. To get a complete view of the data, the teacher dashboard, and the different platform areas, you can log in and then type one of the following teacher identifiers into the teacher search bar: “Johanna” or “candis”.

These profiles allow you to explore the teacher space, the preconfigured classes, dashboards, error trends, pedagogical suggestions, and selected student conversations. Student names have been changed for demonstration purposes.

You can also explore some test student profiles, such as “Stessy” or “Cleo”, to observe the learner-side experience: Socratic dialogue, error diagnosis, reasoning-based adaptation, and progress tracking.

P.S. All documents include a FR/EN language switch button, allowing us to read the content in either French or English.

Video pitch