Research

A physics-informed, reasoning-first approach to echocardiographic understanding that treats cardiac function as a latent dynamical system rather than a static prediction task. We frame EF/LV estimation within a joint embedding predictive architecture (JEPA), learning world models that align imaging, geometry, and temporal dynamics. Reinforcement and self-predictive objectives are used to enforce physiological consistency, counterfactual stability, and calibrated uncertainty—reducing hallucination and enabling mechanistic reasoning over cardiac state trajectories.

A closed-ended, multimodal evaluation benchmark designed to remain unsaturated as medical AI scales. Inspired by Humanity’s Last Exam, this project curates thousands of expert-level clinical caselets spanning imaging, waveforms, operative context, and longitudinal EHR timelines. The benchmark emphasizes temporal reasoning, calibration, and safety-critical error, with automatic grading, confidence-aware metrics, and private held-out sets to detect overfitting and benchmark hacking.

A proposal for an open, lifespan-spanning reference atlas of the human cardio–metabolic axis—integrating heart, vasculature, liver, and adipose across ancestries and metabolic states. We outline standardized tissue sampling and metadata, multi-omic + spatial profiling (sc/snRNA-seq, ATAC-seq, CITE-seq, spatial transcriptomics, 3D imaging), and the computational foundations needed for cross-organ harmonization: ontology unification, batch-aware integration, and perturbation-aware representation learning to distinguish conserved versus organ-specific cell states. The atlas serves as a shared coordinate system for interpreting disease perturbations (HF, CAD, arrhythmias, diabetes/obesity, MASLD/MASH), refining disease taxonomy, and guiding cell-type-specific therapeutics.