About me

I built and operate CorneaForge, the on-premise ophthalmic data platform behind OphtaFlow AI at Quinze-Vingts Hospital in Paris. My work connects device ingestion, scientific image processing, model serving, and the experiments that build on those systems.

I work across the whole path: extracting usable measurements from device exports, implementing corneal geometry, building interfaces to inspect examinations, and training image models. Operating the platform gives me concrete research questions; conducting the experiments shows me which parts of the infrastructure need to improve.

A platform in operation

CorneaForge’s deployed services ingest MS-39 and Corvis data, process OCT images, compute features, and render examination views. MinIO retains source objects, PostgreSQL holds structured measurements and processing state, and a FastAPI application serves the OphtaFlow AI interface. Dedicated workers run under systemd on the hospital’s infrastructure.

The application supports examination review, annotation, native video inspection, and existing model outputs. Research dataset builders use the same source lineage while defining the population and variables for each study. My architecture article follows these paths through the deployment and explains why I separated storage, background processing, and interactive access.

Experiments and engineering results

For anterior-segment OCT, I implemented five pretraining recipes on a shared index of 3,427,130 B-scans: full-field masked reconstruction, two matched patch-budget variants, diffusion reconstruction, and global–local JEPA. The recorded results show distinct representation geometries: the patch-limited arms concentrate variation, while JEPA spreads it across more directions. I explain the measurements and their interpretation in Reading the geometry of AS-OCT pretraining.

I also implemented and optimized JEPA across two NVIDIA L40S GPUs. Profiling traced a communication stall back to late input preparation. Overlapping that work reduced update time from 10.246 to 8.787 seconds in a matched comparison, with input, gradient, optimizer-state, and resume checks preserving the experiment.

My Corvis research combines classical corneal region segmentation, motion analysis, and an experimental neural video classifier. I built synchronized inspection views and human-review tools to examine the intermediate results. The classifier’s scores are exploratory and uncalibrated; independent performance evaluation remains open.

ML Training Monitor grew out of this work: a public monitoring and profiling workflow for coding agents. OCT-CUDA explores kernel fusion for OCT reconstruction with cuFFTDx, focusing on the cost of moving intermediate values through GPU memory.

From statistics to research engineering

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. I joined Quinze-Vingts in April 2024, initially as an intern, and continued as a Machine Learning Engineer.

That background connects statistical modeling with the systems I build today. I collaborate with clinicians on the measurements and outcomes that define a useful research question. My publications cover surgical outcome prediction and corneal biomechanics; my articles explain the computational work behind the research.

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 louissi.nassim@gmail.com.