Ammar Bunajmah

Artificial Intelligence Specialization · Cohort 2026

AI-Driven Workflow Optimization and Prediction in CMS Glance

Every CMS physics analysis is steered from first draft to publication through a 17-stage editorial workflow in the Glance platform, where reviews stall and queues form. This project builds a system that measures where analyses pile up and predicts how long each one will wait, early enough to act on. The work has three parts: a forensic reconstruction of Oracle workflow logs spanning 2007 to 2026 and covering 3,147 analyses, which recovers moves the transition log never recorded and corrects queue lengths that a naive reading of the log overstates by more than an order of magnitude; classical baselines from queueing theory and time-series forecasting, including Little's Law, exponential smoothing and age-conditional survival curves; and learned models spanning spatiotemporal graph networks, recurrent and attention-based sequence models, and per-analysis duration models. Every learned component is measured against those baselines under rolling-origin backtesting and against a direct estimate of the sampling noise in the targets, and is adopted only where it demonstrably wins. Negative results are reported as results.

Read the final report (PDF)