Louis Vincent

– Starting as a freelance Data Scientist, looking to collaborate with companies that want to turn complex data into clear, actionable insights.

– Topic: Longitudinal data encoding for medical decision support in telecardiology – Developed advanced ML methods (neural networks, generative models, VAEs, Gaussian Processes) to model temporal and incomplete medical data in a semi-supervised setting. – Designed algorithms to reconstruct, predict, and quantify uncertainty in patient time series. – Improved heart-failure prediction performance (ROC-AUC +0.05 to +0.1), enabling the system to reach FDA-grade thresholds. – Co-created Pythae, an open-source Python library for training and benchmarking generative autoencoders. – Published at NeurIPS (top-tier AI conference) and co-invented two patents on novel algorithms for medical data modeling. – Collaborated with medical, engineering, and product teams, translating clinical needs into concrete AI solutions.

– Homework help and private tutoring for high school and undergraduate students. – Simplified explanation of abstract concepts and motivation coaching for students struggling with mathematics.

– Taught key concepts linking theory and practice in computational statistics + supervised practical coding sessions in Python. – Topics: stochastic gradient methods, Bayesian inference, EM & MCMC algorithms and extensions.

– Preparation of a CIFRE Thesis and study of various Data Science side projects

– Co-invented a patented algorithm for early prediction of heart-failure risk from implantable cardiac device (CIED) data. – Developed a semi-supervised time-series classification model explicitly handling missing and irregular data. – Achieved 94% precision and 5% false negatives on 20-day risk prediction of heart failure, exceeding all exploratory objectives. – Confirmed the predictive value of longitudinal cardiac data, validating the company's R&D direction and paving the way for new clinical tools. – Co-supervised by Stephanie Allassonniere at Centre de Recherche des Cordeliers.

– Research subject: Deep video denoising using noise robust optical flow. – Designed and implemented a supervised deep-learning framework for video denoising using optical-flow consistency. – Combined Task-Oriented Flow and VNLNet architectures into a unified CNN model, leveraging motion information to preserve temporal coherence. – Focused on training stability and convergence of highly non-convex networks – tuned optimization schemes, learning-rate schedules, and loss formulations. – Conducted extensive experiments in PyTorch, analyzing quantitative gains (PSNR, SSIM) on synthetic noisy datasets. – Worked autonomously under research supervision, bridging theory and practice in computer-vision model design.

– Conducted research on state-of-the-art methods for high-dimensional black-box optimization, focusing on surrogate-based and Bayesian approaches. – Designed and implemented a new variant of the SEGOMoe algorithm in Python/C, improving robustness and sample efficiency in large-scale search spaces. – Benchmarked the method against established baselines, analyzing convergence speed and numerical stability. – Strengthened expertise in meta-model optimization, Gaussian processes, and gradient-free search algorithms.