About me

At Quinze-Vingts Hospital in Paris, I build experiments and the infrastructure around them. Some days that means studying representations of OCT images. Other days it means understanding a device export, reconstructing a corneal surface, or finding out why a training run spends so much time waiting.

I like the connection between those problems. A model depends on a dataset; a dataset depends on measurements; measurements depend on software and assumptions that are easy to overlook. I want to be able to follow a result back through that chain.

From statistics to research systems

I came to machine learning through econometrics: a bachelor’s degree in Economics at Paris-Est Créteil, followed by a master’s in Econometrics and Statistics at the University of Angers. That background shaped how I think about experiments. A good fit is useful, but so is understanding what the data can actually tell us.

I joined Quinze-Vingts in April 2024, initially as an intern, and continued as a Machine Learning Engineer. My work spans statistical modeling, ophthalmic device data, corneal feature computation, and research infrastructure. I collaborate with clinicians to connect computational questions to the measurements and outcomes they care about.

What I’m working on

My current representation-learning work focuses on anterior-segment OCT. I adapt and compare masked reconstruction, predictive representation learning, and diffusion-based pretraining. Crop design, representation diagnostics, and keeping patients separated across evaluation splits matter as much as choosing a model.

For Corvis videos, I’m developing a pipeline for corneal region extraction and experimental classification. The current prototype combines classical image processing with a compact neural classifier. I inspect the intermediate regions and frame representations, keep segmentation failures explicit, and group model development by patient. Anatomical segmentation accuracy and independent classification performance remain open evaluation questions.

I also build CorneaForge, an on-premise platform for ophthalmic data and research workflows. It connects device ingestion, geometry and feature computation, annotation, and dataset construction. The OphtaFlow AI interface makes examinations, derived measurements, and native videos inspectable. A recurring design question is how to preserve the link between a derived value and the source that produced it.

Recent systems work includes a two-GPU JEPA training implementation and profiling the input preparation that feeds it. An apparent communication bottleneck led back to work arriving late from the coordinator. Overlapping that preparation reduced update time; gradient, optimizer-state, and resume checks helped establish that the execution change preserved the intended experiment.

On the systems side, I profile training and explore lower-level implementations when they help explain a bottleneck. That work led to ML Training Monitor, a public monitoring and profiling workflow for coding agents. OCT-CUDA explores a related question: how keeping intermediate values on the GPU chip changes the cost of an OCT reconstruction pipeline.

How I approach the work

Follow a result back to its data. Make the comparison fair. Measure where time goes. Check what changed after an optimization.

My articles explain that process through concrete examples, including the parts that are easy to get wrong. My publications cover research on surgical outcome prediction and corneal biomechanics.

Get in touch

If you’re working on ophthalmic ML, scientific computing, or the systems behind research, I’d be happy to compare notes. You can reach me at [email protected].