Inference-Time Adaptation of Vision-Language Models State-of-the-art few-shot detection, zero-shot classification, and alignment with frozen, black-box VLMs. When Can Models Trust Their Training Data? Diagnosing when and why multimodal models generalize, and turning the diagnosis into fixes. Interpretable Lung Ultrasound Biomarkers Clinician-understandable biomarkers that decouple feature learning from downstream clinical tasks. Learning with Label Structure and Uncertainty Getting more out of imperfect labels, from coarse annotations to class taxonomies to RF waveforms. Patents & Translational Work Four U.S. patents translating the research into deployable clinical ML systems.