I am an MSc Data Science student at EPFL, working at the intersection of computer vision, large-scale multimodal learning, medical imaging, and foundation models. My current work at VILAB focuses on building scalable image and video datasets using distributed pseudolabeling and tokenization pipelines for downstream visual learning.
Previously, I completed a double BSc in Data Science and Engineering and Engineering Physics at the Universitat Politècnica de Catalunya through CFIS, UPC's selective double-degree engineering program. For my bachelor's thesis, I completed a year-long research internship at the NYU Courant Institute, supervised by Prof. Carlos Fernández-Granda, on foundation models for brain MRI.
My main interest is representation learning: how visual models build internal abstractions of geometry, motion, and semantics from raw data, and how these representations can generalize beyond the exact distribution they were trained on.
Building a large-scale multimodal image and video dataset and benchmarking state-of-the-art pseudolabelers for optical flow, depth, surface normals, self-supervised features, and related visual signals.
Developing distributed GPU-accelerated pipelines to pseudolabel and tokenize large-scale image and video data efficiently.
Collaborating with Stanford's spin-off Metamorphic
VILABBuilt a controllable text-to-motion pipeline for generating human motion from language and user-specified paths.
Decoupled global root-path control from local body motion using a path controller and caption-conditioned body GPT, supporting online adaptation to new text prompts or path constraints during generation.
Adapted SAM 2 to 3D CT volumes using inference-time memory and propagation strategies, and evaluated the method on 2,500 CT scans from TotalSegmentator.
Achieved up to a 12.5% Dice improvement on bone structures.
Biomedical Imaging GroupBuilt a large-scale brain MRI dataset with anatomical and clinical labels, pretrained ViT models with MAE and DINO, and designed a supervised segmentation pretraining stage to add anatomical priors to image embeddings.
Bachelor thesis supervised by Prof. Carlos Fernández-Granda. The thesis is available here, and the related paper is available on arXiv.
NYU CourantGPA: 6/6. Top grades in Markov Chains and Algorithmic Applications, Mathematics of Data, and Machine Learning.
GPA: 8.4/10 and 8.6/10. Selected for CFIS, UPC's highly selective double-degree engineering program.