Greater Paris Metropolitan Region
ML Research Engineer with a decade building and deploying ML systems at production scale, and a deep recent focus on generative AI. I work across the full arc, from state-of-the-art research to systems running in front of millions of users. ICCV 2025 Oral co-author and technical lead, with hands-on expertise in Latent Diffusion Models, flow-based architectures, 3D Gaussian Splatting, and self-supervised foundation models. The common thread: shipped to real users, not just benchmarks. I've led cross-functional research from prototype to global deployment, partnered with industrial stakeholders, and translated cutting-edge papers into production-grade systems. Research interests: generative models (diffusion, flow matching, score-based), 3D scene representations, physics-informed ML, AI-accelerated simulation, digital twins, and multimodal models. Outside work, I build and open-source from scratch, diffusion models, a chess Transformer (~2100 Elo), and a Rust path tracer, and write about applied ML.
Leading generative AI, 3D, and physics-informed ML research from prototype to global production. Generative Models & 3D Scene Representations - Designed and trained a custom Latent Diffusion Model for spatially controllable face aging with efficient spatial conditioning — enables real-time generation on iPad Pro. (ICCV 2025 P13N Workshop, Oral) - Developed a self-supervised foundation model for semantically rich facial representations, enabling zero-shot transfer across downstream tasks: attribute recognition, skin segmentation, and age estimation. - Built a novel 2D/3D Gaussian Splatting + FLAME pipeline for high-fidelity facial capture and reconstruction, combining a physics-based parametric face model with learnable Gaussian primitives, optimised for production deployment. Physics-Informed ML & Surrogate Modelling - Deployed low-latency ViT/CNN-based skin tone detection and optical shade matching (eShadefinder) globally across 9 L'Oréal brands on mobile and web, serving millions of users. - Designed and deployed Kubelka–Munk physics-based ML surrogate models for targeted colour formulation across lipstick, hair colour, and foundation. Coupled optical physics equations with neural predictors to replace costly iterative lab experiments with real-time, ML-guided inverse design — validated against industry colour-science standards. Data Pipelines, Production & MLOps - Built and maintained large-scale automated data generation and curation pipelines for model training and evaluation across multiple product lines. - Architected a high-performance model serving pipeline (FastAPI, Docker, ONNX) achieving sub-100 ms inference for real-time product recommendations at global scale.
- Designed an ML-powered self-calibration system for a formulation robot using Gaussian Process Regression with active learning and iterative surrogate modelling to minimise experimental campaign cost. Reduced material waste by 71% and machine downtime by 83%. - Restructured the embedded software team and led safety compliance certification for robotic systems under strict industrial constraints.
- Engineered a real-time fall and tilt detection system for UAVs using multi-sensor fusion (Kalman filtering) for fault-tolerant flight control. - Boosted UAV propulsion efficiency by 44% and cut power usage by 10% through ML-based optimisation and mechanistic aerodynamic modelling, an early application of surrogate-driven system optimisation.