Projects

Predict missing product attributes by modeling the catalog as a graph of products, categories, and interactions. Graph neural networks use these relationships to improve accuracy without relying on heavy text-only models.

Face recognition models are often opaque. We build explanations that reveal which facial cues and conditions drive predictions to improve robustness and reduce bias in biometric systems.

We use deep learning to decode EEG/fMRI signals and map them to intents or commands. This supports assistive interfaces and offers insight into neural processing.

Find specific moments in long videos from text or visual queries. We focus on efficient models that scale to large archives and stay robust to motion, occlusion, and lighting changes.