Computer vision and machine learning researcher
I'm currently at an early-stage medtech startup; before that I was a postdoc in the group of Thomas Demeester. My PhD was on the intersection of visual perception, (stochastic) dynamical systems and physics priors. I started ongoing collaborations with Tolga Birdal and Manfred Opper on SDEs driven by fractional Brownian motion, leading to an ICLR spotlight, and subsequent work led by Gabriel Nobis on generative fractional diffusion models and fractional Schrödinger bridges.
During my PhD I co-founded Hippo Dx, a medical-device start-up behind the automated skin prick test (SPAT), where I worked on the camera and lighting system, image processing and AI-assisted diagnostics. Before my PhD I worked as innovation engineer at CNH Industrial on combine-harvester automation, and as AI lead at Octinion on agricultural computer vision projects.
I live in Brugge with my wife and three kids.
Full list on Google Scholar.
Advances in Neural Information Processing Systems (NeurIPS), 2025
A generative diffusion-bridge framework driven by an approximation of non-Markovian fractional Brownian motion, capturing the memory and long-range dependence that standard Brownian-motion bridges miss.
Machine Vision and Applications, 2025
A three-stage method for 6D pose estimation of texture-less industrial objects: predicting center and curvature heatmaps with a fully convolutional network, triangulating 3D object centers across views, then refining poses by render-and-compare; it reaches millimeter-level, state-of-the-art accuracy on T-LESS and is deployed in a real robotic pick-and-place system.
Nature Communications, 2025
An AI-assisted readout method that measures wheal diameters from Skin Prick Automated Test (SPAT) images, validated against physician measurements, to support consistent allergy diagnosis.
Advances in Neural Information Processing Systems (NeurIPS), 2024
The first continuous-time score-based generative model driven by fractional diffusion: replacing Brownian motion with fractional Brownian motion (Hurst index H) for better diversity and convergence.
The Twelfth International Conference on Learning Representations (ICLR), 2024
The first variational inference framework for non-Markovian neural SDEs driven by fractional Brownian motion, approximating the fBM by a linear combination of Markov processes; derives the variational posterior and ELBO.
Neurocomputing, 2024
Learns Lagrangian dynamics from images: keypoints give Cartesian coordinates with explicit holonomic constraints, trained unsupervised end-to-end, enabling long-term video prediction and energy-based control.
Differential Geometry meets Deep Learning (DiffGeo4DL NeurIPS Workshop), 2020
An unsupervised method to learn the orientation of symmetric objects in images, mapping in-plane rotations into an autoencoder's latent space via a crossing loss, disentangling orientation from other factors.
Interactive prime spiral.
A 2-D fractional Brownian motion (fBM) field rendered live in the browser. Drag to pan, scroll to zoom. Zooming and panning are unbounded, you can keep going indefinitely. Since fBM is self-similar (fractal), the statistics of the noise are the same at every scale, and while zooming, the value under your cursor never changes as details emerge around it. Panning, however, can shift the value of the noise, so as you wander, you eventually drift into other parts of the color palette. Explore this infinite noise space and look for interesting structures, colors and textures. Do not forget to tune the roughness with the Hurst index H, a fundamental parameter governing the structure of fBM.
An interactive talk that builds fractional Brownian motion up from first principles, with live in-browser demos. My talk for the Ghent Workshop on Machine Learning and Biology (2026).
A practical, hands-on introduction to stochastic differential equations (SDEs), with self-contained example code in JAX and Diffrax.
Runs entirely in your browser (all inference in the frontend) and gives live feedback on your sitting posture. We made this during COVID for the Full Stack Deep Learning course.