Establishing a sovereign infrastructure for higher education built on metacognitive student learning profiles.

Team name: UTC EduTech

Use of AI tools :

• Open Source LLMs families: Gemma, Llama, Mistral

  • LLMs are used within our application to answer student questions, either as user-facing LLM, or as integrated subagents.

• We also used private LLMs to support the development of the application (coding assistants as Claude Code). All outputs were reviewed.

External feedback & contributions : Name – Role – Type of contribution (feedback, comments, mentoring, discussion, keynotes, etc.)

People who supported UTC EduTech during the AI Grand Challenge 2026 process:

• Matthieu Bricogne-Cuignières - Associate Professor at Université de Technologie de Compiègne - Guided us from the start, presented us the UTC AI servers and how to use them, proposed resources to guide our reasoning.

• Stéphane Poinsart - Instructional Support Specialist, Content Engineering Unit (ICS) at Université de Technologie de Compiègne - Guided us on examples and existing AI platforms.

• Alain Goudey - Associate Dean for Digital, NEOMA Business School - Keynote presentation on AI usage, European skepticism, and legitimacy of generative AI in higher education

• Alexandra Pregent - CEA Saclay - Keynote on AI ethics, ALTAI framework for trustworthy AI, and harm foreseeability

• Bastien Guerry - Former developer at DINUM, open source maintainer - Keynote on open source licensing, AI sovereignty, and legal status of AI weights

• Caroline Beslin - Project methodology expert, Ecole Centrale de Lyon - Keynote on four-dimension project consolidation framework (Intellectual, Material, Social, Emotional)

• Justine Cassell - Senior researcher at Inria, on-leave Dean's Professor at Carnegie Mellon University - Community comment recommending examination of CMU's Open Learning Initiative (OLI)

• Mahdi Ayadi - Fellow AI Grand Challenge participant - Community comment on student behaviour and the need for structural rather than voluntary solutions

• Chinmay Das - Fellow AI Grand Challenge participant - Community comment on learning profile design and avoiding reinforcement of incorrect assumptions

• Nipun Ranchhod Navadia - Fellow AI Grand Challenge participant - Community comment affirming the sovereign, student-owned learning profile concept and GDPR framing

Initial contribution: Establishing a sovereign infrastructure for higher education built on metacognitive student learning profiles.

Final contribution:

Paper: https://drive.google.com/file/d/1dgdzOr-DXTSeDvOI3nMEaWISuAxj4-2J/view

Website: https://alexandre.amrani.fr/utc_edutech_ai_challenge/

Reflection on the process (important)

We explain here how our contribution evolved during Phase 2:

Comparison between initial publication (13 April 2026) and final main.tex

--- Summary of Evolution ---

The initial submission was a conceptual proposal describing an intended system.
The final submission is a complete academic paper documenting a working minimum
viable product with grounded theoretical work, explicit engagement with
expert input, and transparent documentation of internal debates.

The transformation occurred across five dimensions: specificity, theoretical
grounding, debate transparency, community integration, and infrastructure depth.

--- 1. Key feedback that influenced the work ---

A) Keynote presentations (directly integrated into the argument structure)

- Alain Goudey's keynote introduced the concept of "European specification" for
  AI (privacy-respecting, sustainable, cognitively sovereign). This reframed
  what was previously a technical constraint (using open-source models) into a
  deliberate design philosophy aligned with European values. The final paper
  dedicates substantial argument to this framing, which was absent from the
  initial submission.

- Alexandra Pregent's keynote on the ALTAI framework and harm foreseeability
  gave us a structured vocabulary for ethical design. Her distinction
  between prevention and post-assessment directly shaped the final paper's
  discussion of paternalism and the graduated response structure (default,
  redirection, scaffolding modes).

- Bastien Guerry's analysis of open source licensing and AI weights provided
  the legal-philosophical foundation for the sovereignty argument. The initial
  submission mentioned ILaaS as a technical option; the final paper positions
  sovereignty as inseparable from pedagogy, citing Guerry's argument that a
  model that is merely accessible rather than genuinely open does not confer
  sovereignty.

- Caroline Beslin's four-dimension framework (Intellectual, Material, Social,
  Emotional) revealed that the team was over-investing in the technical
  dimensions and neglecting the social/emotional ones. This directly led to
  features like the transparent, editable profile dashboard that makes the
  student feel seen rather than surveilled.

B) Community comments (explicitly addressed in Section III)

- Mahdi Ayadi (29 April) argued that students knowingly act counterproductively
  and that a "better tool" won't fix a behavioural problem. This pushed the
  team to make the thinking-first interaction the default, non-toggleable
  pattern rather than an optional mode. The final paper acknowledges this
  explicitly and structures the response around structural rather than
  voluntarist solutions.

- Chinmay Das (21 April) asked how the learning profile would avoid reinforcing
  incorrect assumptions. This led to the hybrid model where AI-generated
  inferences are presented transparently to the student, who can annotate and
  correct them. The final paper dedicates a full paragraph to this design
  choice and cites Das by name.

- Justine Cassell (30 April) noted that learning profiles already underlie
  successful AI tutors and recommended studying Carnegie Mellon's Open
  Learning Initiative (OLI). This challenge strengthened the proposal by
  forcing the team to articulate what was already there and how they
  differ from it. The final paper
  explicitly states the differentiation: OLI does not build persistent
  profiles that travel with the student across courses and institutions,
  nor does it give students control over their data.

- Nipun Ranchhod Navadia (10 May) affirmed that the sovereign, student-owned
  learning profile concept is "strong and timely" and highlighted the
  rebalancing of power between institution and learner. This validated the
  direction and is cited in the final paper's discussion of portability.

--- 2. Specific changes made as a result ---

A) From conceptual to implemented

  Initial: "This proposal outlines a conceptual framework that we intend to
  explore and refine during the project, rather than a fully specified or
  already implemented system."

  Final: Describes a working system with FastAPI backend, React frontend,
  Supabase PostgreSQL, multi-agent orchestration, agentic RAG architecture,
  and three-state interaction classifier. The paper includes a dedicated
  Section V ("What Was Built") with full implementation details.

B) From "thinking-first" to "metacognitive" framing

  Initial: Used "thinking-first" as the core concept, with a single interaction
  pattern of "ask before answer."

  Final: Adopts "metacognitive" as the precise learning-science term. The
  interaction model is now three distinct modes (default Socratic inquiry,
  redirection when seeking shortcuts, scaffolding when genuinely stuck) with
  state detection based on lexical patterns and conversation history.

C) From vague profile to structured 7-dimension model

  Initial: Mentioned a "learning profile" capturing "reasoning style, recurring
  difficulties, preferred exercise formats, and progress over time."

  Final: Specifies exactly seven dimensions: dominant reasoning style, recurring
  misconceptions, demonstrated competencies, preferred exercise format, proximal
  zone indicators, engagement signals, and recent progress. Each is grounded in
  specific learning science frameworks (Bloom's taxonomy, Zimmerman's SRL model).

D) Profile transparency became a feature, not just a principle

  Initial: "A profile the student cannot see, edit, or export is a liability."

  Final: Implements a dedicated dashboard where students read, annotate, and
  correct the system's inferences. Instructors only access anonymised aggregate
  data. This design directly addresses Chinmay Das's concern about silently
  reinforcing incorrect assumptions.

E) Addition of a "direct answer" mode

  Initial: Framed the system as purely thinking-first, with no alternative.

  Final: Pragmatically adds an option to send messages to a direct chatbot
  with no Socratic discourse, acknowledging that students will otherwise go
  to ChatGPT anyway. This was a direct response to Ayadi's behavioural argument.

F) Debates section was created from scratch

  Initial: Had one short paragraph about "the central tension we debated
  between guidance and autonomy."

  Final: Section III is a full argument documenting four major debates
  (equity vs friction, sovereignty vs capability, paternalism, profile
  control) plus the Cassell challenge. Each debate presents both positions,
  explains the arbitration, and cites the relevant input that shaped it.

G) References expanded from 9 to 20+

  Initial: 9 references, all cognitive science papers.

  Final: 20+ references including 4 keynote speaker citations, 4 expert
  resource citations, 4 community comment citations, and additional
  academic sources on intelligent tutoring systems (VanLehn, Kulik),
  motivation theory (Deci & Ryan), self-regulated learning (Zimmerman),
  and pedagogical AI research (Favero, Puech).

H) Infrastructure detail

  Initial: "Run on locally-hosted, open-source models through ILaaS."

  Final: Specifies Gemma and Mistral model families, institutional hardware
  at UTC, fallback client architecture with graceful degradation, and
  explicitly addresses the performance trade-off with frontier proprietary
  models, citing Favero et al. for independent validation.

Video pitch

https://drive.google.com/file/d/1zRzLU1gECv2wfYPfPy3K7zn2xVD0dNBv/view