RouteAI

RouteAI is an advanced AI-powered platform designed to revolutionize supply chain visibility by accurately predicting shipment locations, arrival times, and conditions. Utilizing sophisticated generative AI models, RouteAI uniquely integrates geospatial, temporal, and contextual shipment data to forecast exceptions and disruptions in real-time. Unlike traditional models relying on historical averages, RouteAI dynamically adapts to changing conditions, significantly enhancing reliability and operational efficiency in logistics management. By providing precise and forward-looking shipment insights, RouteAI empowers logistics teams to proactively manage routes, minimize disruptions, and optimize resource utilization.

Problem & Motivation

Despite the abundance of logistics tracking tools, most systems today only tell you where a shipment has been—not where it’s going. This retrospective visibility is insufficient for handling exceptions, which affect nearly 1 in 5 shipments. RouteAI was created to address this need. By leveraging generative AI, we aim to provide predictive visibility so shippers, carriers, and 3PLs can make proactive decisions, reduce costly delays, and improve customer satisfaction—especially when disruptions like port strikes or geopolitical instability arise.

Data Source & Data Science Approach

We used AIS (Automatic Identification System) data aggregated from Spire, covering over almost 200k unique tracking sequences across global sea lanes between 2022 and 2025. After intensive preprocessing—including route validation, ghost route removal, and H3-based geospatial indexing—we developed a generative LSTM model. Our architecture encodes both spatial and temporal signals, predicts next-position and next-time jointly, and integrates custom logic to constrain predictions to feasible maritime paths. This allows the model to reason about both where and when a vessel is likely to arrive—far beyond average-case estimates.

Evaluation

We benchmarked our model on three key metrics: GoodRoute accuracy, Mean Absolute Error (MAE) in hours, and day-level timestamp accuracy. Our LSTM-based architecture achieved a 74% GoodRoute match rate, reducing ETA MAE from over 830 hours (baseline) to just ~74 hours. Accuracy within ±1 day improved to over 33%, climbing past 72% within ±4 days. Qualitative tests also showed strong generalization in scenarios such as rerouting around landmasses or navigating narrow canals—critical for operational decision-making.

Key Learnings & Impact

Our biggest takeaway was the importance of contextual encoding—embedding from/to port locations greatly enhanced the model’s understanding of route structure. We also learned that discretizing the geospatial space helped prevent hallucinations common in generative models. RouteAI has the potential to transform maritime logistics by enabling forward-looking scenario planning, real-time rerouting, and exception classification. It lays the foundation for future tools like route simulators and confidence intervals on predicted arrival windows.

Acknowledgements

Special thanks to our mentors, domain experts and the DATSCI 210 faculty. Your insights, support, and feedback helped bring RouteAI to life.