This project builds a streaming analysis pipeline for LCLS imaging data using rank-adaptive matrix sketching to keep memory bounded while preserving structure in high-dimensional images.
The workflow couples Priority Sampling + Frequent Directions with a parallel tree-merge strategy, then applies PCA -> UMAP -> OPTICS to surface interpretable structure and anomalies in real time.
Read the full writeup in the blog post: Matrix Sketching for Online Analysis of LCLS Imaging Datasets.