Help setting up PASS MAJOR in collaboration with medical students, an online platform dedicated to medical exam preparation.Objective: Improved the efficiency of exam preparation by offering an increased volume of tailored and personalized questions through AI.Contribution:• Enhanced LLM (Large Language Model) with RAG (Retrieval Augmented Generation) to create personalized MCQs aligned with exam standards from collection of academic courses and exam MCQs. • Creation of a Chatbot based on academic courses.

Research activities to improve performance of SUPPONOR’s AI-based virtual advertising system:• Research on temporal properties by using several approaches (temporal loss, previous prediction/image, memory module, vision transformer) to improve real-time video segmentation of the advertising spaces.• Research on graphics detection to reduce integration complexity and bandwidth, detect graphics added to the video signal (scores, time, player name, etc.), and ensures that virtual ads are correctly rendered. Work from research to production (Proof of Concept).• Upgrade personal skills in classical computer vision and GenAI (LLM, RAG, JEPA).Other contributions:• Patent monitoring: keep informed about the patent landscape within the industry and identify patents and applications necessitating specific actions (objections, licensing, litigation).• Mentoring a final year of master degree in engineering: recruitment, supervise, write research proposal.• Event participation: VivaTech (coaching on pitch preparation and media training), WAICF, Science Festival (France), PoleSCS conference pitch.

Deep Learning applied to the processing of satellite images.Research and development of computer vision algorithms applied to super-resolution, pansharpening (image fusion), and image restoration processes.Purpose: improve the resolution of satellite images.Technical environment: Python, PyTorch, TensorFlow, Docker

Active Learning for deep neural networks.Research within an international team on active learning. Purpose: define a strategy to train a model from as little annotated data as possible, to increase the robustness of model from adversarial attacks.Supervisor: Frederic Precioso (head of Artificial Intelligence at the French Research Agency).Technical environment: Python, TensorFlow