This project started with a straightforward frustration: watching a regional trucking company run the same inefficient routes for years because no one had time to recalculate them. The goal was to build a working prototype that uses historical delivery data, real-time traffic feeds, and weather patterns to suggest better routing decisions automatically.
What the project actually does
The core system ingests GPS logs from a simulated fleet of 18 vehicles and applies a modified Dijkstra algorithm layered with a gradient boosting model trained on 14 months of delivery records. It does not just find the shortest path — it weighs load capacity, driver shift limits, and fuel consumption per segment.
Testing against the historical dataset showed meaningful reductions in idle time and backtracking, though results varied significantly by region and season. Coastal routes with predictable congestion patterns responded well; rural routes with sparse data were harder to optimize reliably.
Where machine learning fits in
The ML component handles anomaly detection — flagging routes where predicted arrival times consistently drift from actuals. This turned out to be more useful than the routing suggestions themselves, since it surfaces systemic problems a dispatcher might not notice across hundreds of weekly runs.
- Dataset: 14 months of simulated freight logs across 3 provinces
- Model: XGBoost for delay prediction, Dijkstra variant for path selection
- Tools: Python, OpenRouteService API, Tableau for visualization
- Validation: cross-referenced against publicly available Transport Canada delay benchmarks
The hardest part was not the algorithm. It was cleaning 14 months of GPS data where timestamps were off by up to 4 minutes due to device sync errors.
This project is suitable for students interested in applied ML, logistics operations, or data engineering. The full codebase and documentation are available for review.