Projects and Achievements
Since January 2025, our Center has started to prototype and test AI-driven solutions internally at EPITA and EPITECH, with pilots expanding to other IONIS schools. These tools, developed under the joint supervision of educators and instructional engineers, aim to enhance efficiency, personalization, and learning outcomes.
Autonomous Assessment System
Assessment represents one of education's most critical yet resource-intensive activities, with educators often spending countless hours providing meaningful feedback on student work. Traditional grading approaches face inherent challenges: subjective interpretation, inconsistent application of criteria, time constraints that limit feedback depth, and the cognitive fatigue that affects evaluation quality. Our Autonomous Assessment System revolutionizes this landscape by deploying pedagogically-aligned large language models that maintain the nuanced understanding of expert educators while delivering scalable, consistent evaluation.
At its core, the system employs fine-tuned large language models specifically calibrated for educational assessment across diverse academic disciplines. These models undergo extensive training on thousands of expertly-graded assignments, absorbing the subtle pedagogical reasoning that distinguishes exceptional feedback from mechanical evaluation. Unlike generic AI systems, our models understand academic context, recognize common student misconceptions, and apply domain-specific evaluation criteria with the sophistication of experienced educators.
The assessment pipeline integrates a series of analytical layers to ensure comprehensive evaluation. For code projects, static analysis engines examine syntax correctness, algorithmic efficiency, code structure, and adherence to best practices, while the LLM component evaluates creativity, problem-solving approach, and code documentation quality. For written assignments and theses, sophisticated natural language processing techniques assess argumentation quality, evidence integration, clarity of expression, and adherence to academic standards.
What distinguishes our system is its agentic engine, which translates high-level learning objectives into granular, measurable assessment criteria. Educators define course-specific goals and evaluation standards through an intuitive interface, and the AI automatically generates detailed rubrics that maintain consistency across all assignments while allowing for contextual flexibility. The system learns from educator feedback, continuously refining its understanding of quality indicators and assessment priorities for each specific context.
The feedback generation mechanism operates on multiple levels, providing both summative scores and formative guidance. Students receive detailed, personalized comments that identify specific strengths and areas for improvement, along with actionable recommendations for enhancement. The system generates different feedback styles—from encouraging developmental comments for struggling students to challenging extension suggestions for high achievers—ensuring that each learner receives appropriately calibrated guidance.
Quality assurance mechanisms ensure assessment reliability through multi-model validation, confidence scoring, and human oversight protocols. When the system encounters ambiguous cases or detects potential errors, it automatically flags submissions for human review, maintaining the balance between automation efficiency and pedagogical judgment. Extensive bias detection algorithms monitor for fairness across student demographics, ensuring equitable evaluation regardless of writing style, cultural background, or academic preparation level.
Early deployment across IONIS institutions has started to demonstrated encouraging results: significant reduction in grading time while maintaining or improving feedback quality, increased educator satisfaction with assessment timeliness and detail, and enhanced consistency in evaluation standards across different instructors and sections. Perhaps most significantly, educators report being able to redirect their time from mechanical grading tasks toward higher-value activities like curriculum development, individualized mentoring, and innovative teaching strategies.
The system's impact extends beyond efficiency gains to transform the entire assessment paradigm. With automated preliminary grading, educators can implement more frequent low-stakes assessments that provide continuous learning feedback without overwhelming their workload. This shift enables a more responsive, adaptive teaching approach where instruction can be adjusted based on real-time understanding of student progress and common difficulty areas identified through pattern analysis across all submissions.
Dropout Prediction with Conformal Prediction Engine
Student retention remains one of higher education's most pressing challenges. Traditional early warning systems typically rely on reactive indicators—by the time declining grades become apparent, intervention opportunities may already be limited. Our innovative conformal prediction engine revolutionizes this approach by providing proactive, statistically rigorous predictions with quantified uncertainty measures.
Our system employs a multi-modal analysis framework that synthesizes diverse data streams to create comprehensive student profiles. Quantitative behavioral traces—including assignment submission patterns, participation frequency, grade trajectories, and platform engagement metrics—are combined with qualitative insights extracted from student essays and written assignments using state-of-the-art large language models. This LLM-powered textual analysis identifies subtle linguistic markers of disengagement, stress, confusion, or declining motivation that human reviewers might miss.
What distinguishes our approach is the integration of conformal prediction methodology, which provides mathematically guaranteed prediction intervals rather than simple point estimates. This means educators receive not just a risk score, but a statistically valid confidence range—enabling more informed decision-making about intervention strategies. The system can distinguish between students who are definitively at risk, those in uncertain territory requiring closer monitoring, and those who are likely to succeed independently.
The predictive model operates continuously throughout the academic term, updating risk assessments as new data becomes available. Early pilots across IONIS schools have demonstrated the ability to identify at-risk students up to 8 weeks before traditional indicators would surface, with usable accuracy in detecting students who eventually withdraw or struggle significantly. Crucially, the system maintains transparency by providing explanatory features that help educators understand why a particular student has been flagged.
This proactive intelligence transforms student support from reactive crisis management to preventive care. Academic advisors can prioritize their limited time on students most likely to benefit from intervention, while customized support strategies can be developed based on the specific risk factors identified for each individual. The result is a more efficient allocation of institutional resources and, most importantly, improved student outcomes through timely, targeted support that addresses problems before they become insurmountable.
Aligned Content Generator
Traditional assessment creation is one of the most time-consuming aspects of pedagogy, often requiring hours to craft meaningful questions that properly evaluate student understanding while avoiding common pitfalls like obvious answers or misleading distractors. Our Aligned Content Generator addresses this challenge by intelligently automating the creation of high-quality, adaptive assessments.
The system leverages Natural Language Processing approaches to analyze course materials—including lecture notes, textbooks, and supplementary resources—and automatically generate contextually relevant questions. Our AI generates multiple-choice questions (MCQs), true/false statements, and open-ended prompts that are precisely aligned with learning objectives and curriculum standards.
What sets our generator apart is its intelligent distractor creation engine. Rather than producing generic wrong answers, the system identifies common misconceptions and generates plausible but incorrect options that help reveal specific gaps in student understanding. The difficulty adjustment mechanism continuously learns from student performance data, ensuring that each learner receives questions that are challenging yet achievable—maintaining what educational psychologists call the "zone of proximal development."
Seamlessly integrated into our existing Learning Management Systems (LMS) platforms, the tool operates transparently within educators' familiar workflows. Teachers can review, modify, and approve generated content before deployment, maintaining pedagogical control while dramatically reducing preparation time.
The transformative impact extends beyond efficiency gains. Students engage with more frequent, personalized practice opportunities that adapt to their individual learning pace and style. This continuous assessment approach provides immediate feedback loops, helping learners identify knowledge gaps before they become entrenched. Meanwhile, educators gain unprecedented insights into class-wide learning patterns, enabling data-driven instructional decisions and targeted interventions where they're needed most.
Conversational Pedagogical Assistant
In today's fast-paced educational environment, students often encounter learning obstacles outside traditional office hours, leading to frustration and potential disengagement when immediate help isn't available. Our advanced Conversational Pedagogical Assistant addresses this critical gap by providing intelligent, context-aware support that bridges the temporal and spatial limitations of conventional teaching methods.
Built on Retrieval-Augmented Generation architecture, our AI assistant combines the conversational capabilities of large language models with precise, course-specific knowledge retrieval. The system maintains comprehensive, dynamically updated knowledge bases for each program, incorporating lecture materials, textbooks, assignment guidelines, past student questions, and curated educational resources. This ensures that responses are not only accurate but also pedagogically aligned with specific course objectives and institutional standards.
The assistant employs sophisticated natural language understanding to interpret student queries across multiple contexts—from conceptual clarifications and problem-solving guidance to procedural questions about assignments and deadlines. Rather than providing direct answers, the system is designed to guide students through Socratic questioning techniques, encouraging critical thinking and independent problem-solving while offering appropriate scaffolding when students become stuck.
Currently in pilot testing across multiple IONIS programs, the assistant operates continuously in secure test environments, providing 24/7 accessibility that accommodates diverse student schedules and learning preferences. The system tracks interaction patterns and frequently asked questions, generating valuable analytics that help educators identify common misconceptions, curriculum gaps, and areas where additional instructional support may be needed.
Beyond immediate student support, the assistant serves as an intelligent teaching aide that amplifies educator effectiveness. By analyzing conversation logs and student interaction patterns, it surfaces emerging trends in student confusion, identifies topics that may need reinforcement, and provides data-driven insights that inform curriculum improvements and teaching strategies. This creates a continuous feedback loop between AI-assisted learning and human pedagogical expertise.
Early results from pilot programs demonstrate significant improvements in student satisfaction and engagement, with 24/7 availability leading to more consistent learning momentum and reduced anxiety about getting stuck on problems outside class hours. For educators, the assistant provides unprecedented visibility into student learning processes, enabling more targeted and effective classroom interventions.
Domain-Specific Applications
- Computer Science (EPITA, Epitech): Automates code project grading, generates unit tests, detects code similarities, and suggests improvements. This can streamline the evaluation process and help students better understand their coding practices.
- General Engineering (ESME, IPSA): Evaluates technical reports, simulates interviews, generates practical exercises, and analyzes multidisciplinary projects. These applications aim to support both technical skill development and interdisciplinary learning.