Studied Kolmogorov-Arnold Networks as an alternative architectural prior for implicit neural representations of 3D scientific volumes at Argonne National Laboratory. Achieved +2–15 dB PSNR over SIREN/MLP baselines and characterized regimes where learned activation functions outperform fixed-activation INRs via systematic ablation. Trained on the ALCF Sophia DGX A100 cluster.
Delivered sit-to-stand repetition counts with Mean Squared Error < 1 repetition vs. clinician annotations across ~50 sessions. Compared end-to-end vs. keypoint-based architectures for action segmentation from monocular video. Implemented detection and tracking pipelines with MediaPipe, OpenCV, and OpenMMLab models. Funded by the National Cancer Institute.
Engineered a large-scale 1M+ multimodal dataset from missing person cases. Prototyped location predictors comparing transformer-based fusion of vision and tabular features against GBDT and agentic baselines. Sponsored by the Menlo Park Fire Protection District.