I am a Postdoctoral Researcher at KU Leuven in Belgium. My research focuses on understanding and improving multimodal foundation models by analyzing how they integrate information across modalities and identifying modality biases, modality integration failures, and limitations in state-of-the-art vision-language and audio-visual systems.
Previously, I obtained my PhD at KU Leuven, where I worked on uncertainty calibration of deep neural networks in the ESAT-PSI lab, supervised by Prof. Matthew B. Blaschko.
Our workshop Towards Trustworthy Predictions: Theory and Applications of Calibration for Modern AI is accepted at AISTATS 2026 in Morocco.
Preprint for CLASH: A Benchmark for Cross-Modal Contradiction Detection is available on arXiv.
Gave a guest lecture on uncertainty calibration in AI at Aalto University, Finland.
Our paper DAVE on audio-visual evaluation is accepted at NeurIPS 2025 🎉.
Started a new position as Postdoctoral Researcher at KU Leuven.
Successfully defended my PhD (thesis available here) 🎓.
LaSCal, our paper on label-shift calibration without target labels, is accepted at NeurIPS 2024 🎉.
Our paper on estimating proper calibration errors is accepted to AISTATS 2024.
Presented our workshop paper on assessing calibration under covariate shift at the WiCV workshop at ICCV 2023. Served as a student volunteer at ICCV 2023.
Our paper on calibration in object detection is accepted to WACV 2024.
Presented our work on estimating calibration error at NeurIPS 2022.
Recognized as an outstanding reviewer (top 10%) at ICML 2022.
Achieved 1st place on the T1A vector task at the LeQua 2022 challenge.
Featured in Computer Vision News for work presented at MICCAI 2021.
My research focuses on understanding and improving multimodal foundation models, with a particular interest in designing diagnostic benchmarks and analyzing modality integration failures. Previously, my work centered on uncertainty calibration, where I developed theoretically grounded methods for estimating calibration errors and ensuring model reliability under distribution shifts. For a complete list, please visit my Google Scholar.
A diagnostic benchmark for systematically evaluating audio-visual models, preventing visual bias by decoupling evaluation into atomic subcategories. It ensures both modalities are necessary for correct answers.
I develop diagnostic benchmarks and evaluation methodologies for multimodal learning and design techniques to improve model robustness, calibration, and real-world reliability.
Developed and maintained full-stack web applications. Started with an internship Dec 2016 - Feb 2017.
Technologies: Java (Spring, GWT), JavaScript (React), SQL (PostgreSQL, MySQL), Git, Jenkins.
Thesis: Advancing Calibration in Deep Learning: Theory, Methods, and Applications
Thesis: Human-initiated Interactive Learning with Clustering-based Global explanations
Thesis:Comparison of Machine Learning Models for a Job Recommendation System