Beamline Deployment

For a ptycho-tomography dataset of 74 projections, ML-augmented reconstruction reduced per-projection runtime from about 153 seconds to about 41 seconds under standard operating conditions. Containerized reconstruction allows different algorithm versions and dependencies to coexist. Deployment uses Prefect-based GPU-aware orchestration to manage GPU availability, memory, data ingestion, conversion, and reconstruction pipelines. The method has been deployed in production at an ALS beamline inside CDTools.