Blog
Writing, research notes, and projects.
A combined index of technical writeups, project pages, and research notes.
Latent Diffusion Priors for Physics-Based Inverse Problems
Derives DDPM training and applies latent diffusion priors to Bayesian inversion, comparing BBVI and Metropolis MCMC.
Flash-SD-KDE: Accelerating SD-KDE with Tensor Cores
Reorders SD-KDE to expose GEMMs and accelerate density estimation on GPUs.
Matrix Sketching for Online Analysis of LCLS Imaging Datasets
Rank-adaptive matrix sketching plus PCA -> UMAP -> OPTICS for online monitoring of LCLS imaging data.
Score-Debiased Kernel Density Estimation (SD-KDE)
Shift-then-smooth KDE that uses score estimates to reduce leading-order bias.
Unsupervised Learning for Anomalous LZ Waveform Data
Fourier subsampling -> UMAP -> HDBSCAN for clustering LZ S2 waveforms and isolating anomalous populations.
Flash-SD-KDE on Tensor Cores
Reorders SD-KDE to expose GEMMs and accelerate density estimation on GPUs.
Latent Diffusion Priors for Inverse Problems
Latent diffusion models as Bayesian priors for inverse heat conduction; BBVI vs Metropolis MCMC.
Unsupervised LZ Waveform Anomaly Detection
Fourier subsampling -> UMAP -> HDBSCAN to cluster LZ S2 waveforms and isolate anomalous populations.
Matrix Sketching for Online LCLS Monitoring
Rank-adaptive matrix sketching plus PCA -> UMAP -> OPTICS for streaming LCLS imaging data.
Score-Debiased Kernel Density Estimation
Shift-then-smooth KDE that uses score estimates to remove the leading-order bias term.
Welcome
What I'm hoping to write about here — AI-for-Science, interpretability, and building robust systems.