Research notes
Reading the geometry of AS-OCT pretraining
Measured representation changes across five AS-OCT pretraining runs, and the profiling work that cut two-GPU JEPA update time by 14.2%.
Nassim
Louissi
Research Engineer
I build and operate clinical AI systems that bring ophthalmic data and diagnostic predictions to physicians.
At Quinze-Vingts Hospital in Paris, I built and operate OphtaFlow AI: diagnostic support running in production, available 24/7 and in beta testing with physicians. CorneaForge powers its data, automatic predictions and research infrastructure.
Corneal geometrySchematic
to diagnostic features.
Clinical AI in production
Diagnostic results are available in under one minute after examination completion. The service calculates missing predictions automatically and retrieves them when physicians open a patient record.
Research experiments
Five pretraining recipes on 3.43 million AS-OCT images, plus Corvis motion analysis and neural video classification.
Measured performance
Two-GPU JEPA updates reduced from 10.246 to 8.787 seconds, with the training recipe and resume behavior preserved.
CorneaForge / OphtaFlow AI
From device measurements to diagnostic support.
I built CorneaForge to connect ophthalmic devices to the diagnostic workflow in OphtaFlow AI. Its production services ingest MS-39 and Corvis data, process OCT images, compute corneal features, and render examinations for physicians.
When an examination's measurements are ready, the background service calculates missing predictions and stores them. Opening a patient record retrieves the completed results alongside the measurements, images and videos. The same infrastructure supplies research datasets, with each derived result linked back to its device export.
Read the architecture and engineering decisions- Data
- MinIO for source objects; PostgreSQL for structured measurements and processing state.
- Processing
- Dedicated ingestion, feature, prediction, rendering, and OCT workers supervised by systemd.
- Application
- OphtaFlow AI provides physicians with diagnostic support, examination review, and annotation through a FastAPI application.
Building toward January 2027
My next step for OphtaFlow AI: bring corneal and retinal information into one clinical decision platform.
I am working toward three targets for this next stage:
- 100+ diseases
- Diagnostic coverage across corneal and retinal conditions.
- 5 devices
- Combine complementary measurements and images through multimodal fusion.
- 1M+ examinations
- Processed through the platform.
JEPA as the first foundation
Self-supervised learning lets models learn visual structure from the images themselves. I am prioritizing JEPA as the first approach for the next stage, then building toward combining what each device reveals about the eye.
How I train these representationsDiagnosis is one task
My roadmap also includes treatment and surgery recommendations, surgical planning with calculated and prefilled parameters, and estimates of the likelihood of success with their uncertainty. The aim is to connect examination findings to the decisions that follow.
Read the clinical and research roadmapWriting from the work
The mathematics, engineering decisions, and experiments behind the systems I build.
Representation learning
Learning without labels: what MAE and JEPA actually predict
The learning objectives I use for AS-OCT: reconstruct hidden pixels, align views of an image, and preserve variation between sources.
Experimental design
Five OCT pretraining recipes, and the choices behind them
The five AS-OCT recipes I implemented on 3.43 million images: their inputs, objectives, shared controls, and compute tradeoffs.
Clinical AI systems
From device data to real-time diagnostic support
How I built OphtaFlow AI for diagnostic support in production 24/7, powered by CorneaForge, and the next stage of its clinical and research platform.
Scientific computing
From measured points to Zernike coefficients
How I turn native corneal samples into maps and Zernike descriptors in CorneaForge, with explicit reference surfaces, coordinates and units.
Performance engineering
Cache the geometry, then change the values
How I accelerate CorneaForge’s geometry engine with reusable interpolation, batched surface fits, and directly evaluated derivatives.
Scientific computing
From corneal shape to optical aberrations
How I built the optical calculation in CorneaForge: surface slopes, refraction, ray intersections and optical path differences become Zernike descriptors.
Scientific computing
The reference surface changes the residual
How I implemented conoid and biconic fitting in CorneaForge, and why choosing a reference determines which parts of corneal shape remain in the residual.
Open source
Tools and experiments that grew out of questions I needed to answer.
Monitoring & profiling
ML Training Monitor
The monitoring and profiling workflow I use in my own research, packaged as a skill for coding agents.
I used it to trace coordinator stalls in two-GPU JEPA, overlap input preparation with GPU work, and measure a 14.2% reduction in update time. Read the profiling result.
View the repository- Observe the run
- Bring training metrics and GPU activity into the same view.
- Find the bottleneck
- Use profiles to investigate where time and memory go.
- Check the change
- Compare measurements and numerical behavior after an optimization.
OCT-CUDA
A CUDA experiment in fusing an OCT reconstruction pipeline with cuFFTDx, exploring the cost of moving intermediate values through GPU memory.
Selected publications
Machine Learning Model for Predicting Visual Acuity Improvement After Intrastromal Corneal Ring Surgery in Patients With Keratoconus
Perez E*, Louissi N*, et al.
CorvisST biomechanical indices in the diagnosis of corneal stromal and endothelial disorders: an artificial intelligence-based comparative study
Borderie VM, Georgeon C, Louissi N, et al.
Have a research question in common?
I’m happy to talk about ophthalmic ML, scientific computing, and the practical work behind an experiment.
louissi.nassim@gmail.com