My dissertation and applied research spans probabilistic machine learning, astronomical image analysis, and high-performance CT reconstruction, united by a focus on reliability, uncertainty, and scalability in high-dimensional data.
Developed a scalable Bayesian machine learning framework that improves the reliability of predictive models by addressing a key limitation of standard variational inference: the systematic underestimation of uncertainty.
The Trustworthy Variational Bayesian (TVB) approach for logistic regression remains robust under model misspecification while maintaining computational efficiency. By combining principled statistical modeling with high-performance optimization, the method produces better-calibrated uncertainty estimates and more stable predictions in high-dimensional settings.
Designed a nonparametric statistical framework for reconstructing faint astronomical signals by jointly analyzing multi-wavelength imaging data. The approach leverages cross-band correlations between continuum and emission-line observations to improve background estimation and enhance detection in low signal-to-noise conditions.
By integrating spatial modeling, local covariance estimation, and controlled signal-to-noise experiments, this work demonstrates significant improvements over traditional single-band methods.
Developed high-performance data pipelines and GPU-accelerated algorithms for large-scale computed tomography reconstruction at Sandia National Laboratories. This work focused on building end-to-end systems for ingesting, processing, and analyzing high-dimensional imaging data with an emphasis on performance, reliability, and reproducibility.
Through systematic validation, benchmarking, and anomaly detection, the pipeline enables efficient processing of complex datasets while identifying failure modes in real-world inspection systems.