The Problem
A logistics operator managing a fleet of over 500 vehicles across 20+ regions was planning delivery routes manually. Dispatchers assigned routes based on experience and rough geographic intuition — a method that worked when the fleet was smaller and delivery volumes were predictable, but had not scaled to the company’s current size.
The consequences were measurable and compounding. Inefficient routes were costing real fuel. Missed delivery windows were generating customer complaints and penalty clauses. The maintenance team was seeing higher-than-expected wear on vehicles routed through suboptimal paths. And the dispatch team was spending hours each morning on planning work that would need to be re-done each afternoon as real-time conditions changed the optimal routes.
The Constraints
Real-time conditions, not just static optimization. Route optimization that ignores live traffic conditions or weather events produces plans that are theoretically optimal at 6 AM and operationally wrong by 9 AM. The system needed to incorporate live conditions and adjust routes dynamically during execution, not just at planning time.
Fleet diversity. The operator ran vehicles with different capacity classes, load configurations, and operational constraints (some vehicles had refrigeration requirements, others had height restrictions for specific zones). The optimization engine had to treat the fleet as heterogeneous, not uniform.
Scalability across regions. 500+ vehicles across 20+ regions meant the system needed to optimize route planning at a scale that made manual override impractical. Exceptions should be the exception, handled by exception — not a routine part of the daily planning workflow.
Our Approach
The optimization engine is built on Google OR-Tools for the core Vehicle Routing Problem (VRP) solver, augmented with TensorFlow and PyTorch models for predictive traffic and congestion modeling.
The static optimization layer runs nightly: given the confirmed delivery schedule, vehicle fleet state, and geographic constraints, OR-Tools solves for globally optimal route assignments across the entire fleet. Genetic algorithms handle the combinatorial complexity of large-scale multi-vehicle, multi-depot routing.
The dynamic adjustment layer runs continuously during the delivery day. Real-time traffic data from Google Maps Platform and OpenStreetMap feeds a predictive model that anticipates congestion 30–60 minutes ahead, triggering route recalculation for affected vehicles before they hit delays. Weather event detection triggers contingency routing for meteorologically sensitive zones.
Driver-facing routes are delivered through a mobile interface (Leaflet.js) with turn-by-turn navigation and real-time updates. The dispatch dashboard provides fleet-wide visibility — each vehicle’s current position, ETA to next stop, and deviation from planned route — enabling dispatchers to intervene on genuine exceptions rather than managing routine variation.
The Outcome
- 25–28% fuel cost reduction across the fleet — the largest single operational cost improvement in the company’s history
- 30% faster deliveries — average improvement of 11 minutes per delivery, compounding across 500+ vehicles daily
- 15% reduction in maintenance costs — optimized routing reduces mileage and wear per delivery cycle
- Manual planning eliminated across all 20+ regions — dispatchers now manage exceptions rather than create routes
Team
Engagement: 5 months, 4 engineers (1 ML, 1 backend, 1 mobile, 1 data).
Stack: Python, Google OR-Tools, TensorFlow, PyTorch, Scikit-learn, PostgreSQL (PostGIS), Leaflet.js, AWS Lambda/EC2, Google Maps Platform, OpenStreetMap