Building ÒCTAVIA

Òc to intelligence. No to dependence

Borde – Da Silva – Lizy Pousson – Goujon | Licence 3 – IRT - 2025/2026

What situation or context are you examining?

The Real-Life Situation in Higher Education We are examining the dual crisis currently facing university students: a severe organizational overload combined with a shifting pedagogical landscape disrupted by generative AI. At the University of Toulouse, for instance, students consistently report that managing overlapping deadlines, course logistics, and syllabus tracking generates a crushing mental load. Simultaneously, they are trying to navigate complex academic materials in an era where AI offers instant, but often superficial, solutions.

What is your critical analysis?

How AI is currently used in this situation: Currently, generative AI is used by students primarily as a reactive "answer machine." Instead of utilizing AI as a study companion, students often use it as a shortcut to bypass academic effort (e.g., copying and pasting assignment prompts to get immediate, final answers).

Issues, tensions, and transformations it creates:

What positions were debated and arbitrated?

During the development of ÒCTAVIA, our team faced several critical debates that shaped our final architecture:

Debate 1: The Role of the Teacher (Top-Down vs. Bottom-Up)

The Debate

Should the system structurally require professors to be the first link in the chain (generating the core pathways) to ensure quality?

The Arbitration

We observed that teaching staff are already overwhelmed and highly polarized regarding AI. Forcing adoption would create administrative friction. We resolved to make ÒCTAVIA strictly a "by students, for students" ecosystem out of the box, while offering an optional "Teacher Account" for proactive professors on a purely voluntary basis.

Debate 2: Model Power vs. Equity and Ecology

The Debate

Should we rely entirely on the most powerful, state-of-the-art LLMs to guarantee the best reasoning, despite high token costs and energy consumption?

The Arbitration

Recognizing the reality of student financial precariousness, we concluded the tool must be highly optimized and accessible. We implemented "Model Routing": using smaller, low-energy models (like Llama/Mistral 7B) for background tasks (e.g., the Planner Agent) and reserving heavier models strictly for deep Socratic reasoning. This trade-off ensures ecological sobriety and equal educational opportunity.

Debate 3: Feature Focus (Content Generation vs. Organization)

The Debate

What makes us truly different from existing tools like NotebookLM?

The Arbitration

We initially drifted toward highlighting our "sexy" features (Safe RAG, quiz generation). However, after confronting peer feedback, we deliberately re-centered our value proposition on the Planner Agent. We realized that reducing the organizational mental load is the necessary prerequisite before any deep learning can occur.

What contributions are you proposing?

Concrete and Actionable Contributions:

The ÒCTAVIA Multi-Agent Ecosystem: A centralized platform where three distinct agents interact proactively: a Planner (to synchronize schedules and manage mental load), a Content Creator (to transform official PDFs into interactive tools), and a Socratic Tutor (hard-coded to refuse direct answers and guide students via maieutics).

The Scaffolding Index: A proposed evaluation metric to measure true active learning, tracking how many hints the Socratic agent must deploy before the student understands a concept, thus proving growing autonomy rather than just delayed answers.

Conditions for Realistic Implementation: