This project addresses real-time wait-time prediction challenges faced by Waitwhile, an intelligent queue management platform that serves heterogeneous queuing environments across industries. We develop gradient-boosted regression models that incorporate engineered service, resource, and industry features while satisfying strict production constraints on latency and model size. Using event-level visit data, we compare squared-error and quantile loss functions and introduce long-wait weighting and segment-specific models at the industry and location levels. The final pipeline reduces mean absolute error by more than 35% relative to a baseline queue-only model and substantially improves the 90th percentile error and long-wait underprediction rates, leading to more reliable wait-time estimates in high-impact scenarios.
Watch the team present this project at 29:30 in the session recording here.
Keywords: queue management, wait-time prediction, long-tail modeling, real-time prediction, gradient boosting, quantile regression
Faculty Advisor
PhD in Theoretical Physics, MS in Computer Science. Currently – Applied Scientist at Amazon. Previous jobs: Computational Scientist at the Argonne National Laboratory, Scientist at LIGO project of California Institute of Technology. I specialize in scientific computing, HPC, machine learning, data analysis.