Block-Wise Differentiable Sinkhorn Attention
A block-wise optimal-transport attention kernel with an exact tail-refinement adjoint, one-reference-tile schedule, and TPU-oriented JAX implementation.
Paper & implementationI like knowing
how things work.
I’m Dylan, an AI engineer at Valeo and an independent researcher. I work on autonomous driving, efficient machine learning, and the infrastructure around it.
I’ll usually read the source code and trace what happens to the data before changing anything. It takes time, but I’d rather understand why something breaks.
Research
All papersSupport Topology and Gradient Mixing in Sinkhorn Layers
How does a sparse transport graph shape the gradients that pass through it?
Paper & interactive figuresNotes & essays
- Tragedy of the Correlated CommonsEducation & game theory
- The Communist ConstraintEconomics & optimization
- Compounding Returns of Genetic CapitalGenetics & long horizons
Selected work
All projectsGenerative data & ML infrastructure
Built and optimized internal ML workflows spanning generative augmentation, trace processing, orchestration, and cross-cluster scheduling for perception workloads.
Personalized cancer vaccine design
Variants → sequence reconstruction → peptide graphs → MHC ranking.
Domain-shift quantification
Estimate policy suitability before a changed environment exposes failure.
On the bookshelf
GoodreadsCurrently reading
The 7-star shelf
From my public Goodreads shelves · Snapshot: 2026-08-25