Vedran Stipetic

Engineered and deployed a photo ranking service that selects primary images from a corpus of 300M+ images, optimizing for clickthrough rate across hotels, restaurants, and attractions. Developed an image quality pipeline including rotation correction and near-duplicate detection to ensure visual diversity across listings. Built image-text matching models for visual search and contributed to multimodal AI systems for image tagging.

Embedded with Qualcomm, developing perception systems for autonomous driving. Built real-time vehicle detection and tracking models optimized for edge deployment. Implemented Active Learning pipeline using Monte Carlo dropout and normalizing flow density esti mation, improving data selection efficiency and model uncertainty quantification. Led company-wide ML competence group, establishing training curriculum for 20+ engineers.

Created state-of-the-art methods for single image dehazing and single image rain removal based on mathematical modeling and unsupervised learning Focused on adapting non-smooth metrics for image degradation to be useable in deep learning, with special attention on reference-free metrics and adapting them for unsupervised learning Authored onejournal paper and four conference papers on various topics in the field of image restora tion Cooperated with an international team of researchers on a paper about high dynamic range image visualization