Experience

Applied AI Scientist

Orange Business

Flux Vision — Orange Research

Generative models for synthetic human-mobility data, from research prototype to production pipeline.

• Research lead on transformer-based generative models that synthesise realistic mobility traces, letting Orange derive value from operator data without exposing real subscriber trajectories.
• Engineered the data pipelines behind it: Apache Airflow orchestrating distributed processing across multiple hosts, parallelised horizontally with Docker and Celery.
• Took models from notebook to production: reproducible training, tracked experiments, containerised serving, MLOps practice throughout.
• Defined and opened a Master’s-level research internship on generative mobility modelling; scoped the subject and supervised the work.
• Work with network, privacy and product stakeholders to turn research output into deployable systems.

PhD Researcher — Deep Learning for Radio Planning

Orange

Doctorate with UTBM — FEMTO-ST Institute

Replaced slow physical simulation with generative models that predict Wi-Fi coverage directly from a floor-plan image, then optimised access-point placement on top.

• Designed IRGAN, a conditional GAN that predicts a full indoor radio map from a single 2D floor plan at 2.5 dB RMSE — fast enough to sit inside an optimisation loop, unlike ray tracing.
• Extended the approach with conditional diffusion models and a physics-informed ResUNet, placing Top-5 of 16+ teams in the ICASSP 2025 Grand Challenge on generalisable path-loss prediction.
• Solved the sim-to-real gap by fine-tuning on crowd-sourced 5 GHz measurements from live home networks (EuCAP 2024).
• Coupled the learned propagation model with metaheuristics — simulated annealing, tabu search, genetic algorithms — for automatic access-point placement under coverage and throughput constraints.
• Filed a patent on incorporating model uncertainty into indoor Wi-Fi optimisation.
• Published 4 international and 3 national papers; shipped a React + Flask demonstrator used to present the work internally.
• Supervised a Master’s thesis on automatic floor-plan vectorisation; taught wireless systems, Linux and C for two years.

Research Intern — Remote Sensing

LISIC, Université du Littoral Côte d’Opale

Deep learning for multispectral / hyperspectral satellite image fusion.

• Designed a fusion architecture from scratch that uses Sentinel-3 hyperspectral data to fill gaps in Sentinel-2 multispectral time series, raising effective temporal resolution.
• Published the result at ICASSP 2022 and GRETSI 2022 as first author.