I design and operate the MLOps infrastructure at Quinze-Vingts Hospital — one of Europe's leading ophthalmology centers — providing clinicians with 24/7 AI-assisted diagnosis. The system integrates heterogeneous data sources (corneal topographers, biomechanical analyzers, clinical records), applies the necessary mathematical transformations to make that data ML-ready, and serves predictions that feed directly into treatment decisions. I'm responsible for the full pipeline: from raw device exports to production-grade model serving.
I came to ML through an unusual path: a Bachelor's in Economics at Paris-Est Créteil, where I got hooked on econometrics and statistical modeling, followed by a Master's in Econometrics and Statistics at the University of Angers. The transition from economic forecasting to production ML systems turned out to be more natural than you'd expect — the core problems (noisy data, small samples, high stakes, real-time constraints) are the same.
Experience
- Architecting and operating a full MLOps infrastructure for 24/7 AI-assisted clinical diagnosis, integrating data from MS-39, Corvis-ST, and clinical databases (~150k exams).
- Built automated data pipelines: polar-to-cartesian coordinate conversion, Zernike polynomial extraction, biomechanical feature engineering — 4,000+ features computed at 5× speed (30 samples/min), with 2–6% deviation from manufacturer software.
- Trained and deployed predictive models (ResNet, ViT, CatBoost) in Python/PyTorch — 97% accuracy on 1,000 samples, 100% classification on 120 surgical cases vs. 75% baseline, saving €70k–120k/year in avoided unnecessary procedures.
- Deploying models with Docker/Kubernetes; collaborating directly with clinicians on validation and production integration.