About

I am a research scientist working on generative models. The thread running through my work is a single idea: when something is expensive to simulate or impossible to collect, learn a model that generates it instead.

That idea started with radio waves. My doctorate — a CIFRE thesis run jointly between Orange Labs and UTBM — began with a practical problem: planning a Wi-Fi deployment means evaluating thousands of possible access-point layouts, and every evaluation traditionally requires a ray-tracing simulation far too slow to sit inside an optimisation loop. I reframed the problem as image generation. Give a conditional GAN a 2D floor plan, get back a full coverage map. What had taken minutes took milliseconds, which made automatic access-point placement tractable for the first time. Later work replaced the GAN with conditional diffusion and a physics-informed ResUNet that generalises to buildings and frequencies it has never seen.

The same idea moved to language, for more personal reasons. Bambara is my mother tongue and the first language of millions of people across Mali and the wider Mande region — yet machine translation for it barely works, because the parallel data does not exist. Kumatigi was my answer: a curation-and-augmentation pipeline that scores every sentence pair for quality, released as an open 67k-pair French↔Bambara corpus and published at Findings of ACL 2026. The finding that mattered was that filtering aggressively for quality beat simply collecting more data — which is the opposite of the usual advice, and useful for anyone working on an under-resourced language.

Today at Orange Research I work on generative models for human mobility: learning to synthesise realistic movement traces so that operator data can be put to work without any real subscriber being exposed. It is the same problem shape again — sequences this time rather than images or sentences.

I taught for two years alongside the doctorate — wireless systems, Linux, C — supervised a Master’s thesis, and now scope and supervise research internships. I review for NeurIPS and AAAI. I am qualified by the CNU in sections 27 and 61, so the academic route stays open, but what I care about is research that reaches something real.