GreenRoute: AI Route Optimization for Last-Mile Delivery
A route optimization platform using ML to reduce delivery times by 28% and fuel costs by 19%. Processes 15,000+ daily deliveries across the GCC region.
28%
Faster deliveries
19%
Fuel cost reduction
96%
On-time rate
15K+
Daily deliveries
The challenge
GreenRoute Logistics handles last-mile delivery for e-commerce companies across Dubai, Riyadh, and Doha. With 15,000+ daily deliveries across 200 drivers, their manual route planning was hitting a wall. Dispatchers spent 3 hours each morning assigning routes based on experience and intuition — and the routes were often suboptimal, leading to late deliveries, high fuel costs, and frustrated drivers.
The problem was compounded by the unique challenges of GCC delivery: extreme heat requiring cold-chain compliance for certain goods, gated communities with restricted access hours, new developments not yet on Google Maps, and traffic patterns that shift dramatically during Ramadan and holiday periods.
Our approach
We built a three-component system: a Python ML service for route optimization, a Flutter mobile app for drivers, and a Next.js web dashboard for dispatchers. The ML service uses a modified vehicle routing problem (VRP) solver enhanced with machine learning for travel time prediction.
The core algorithm combines Google OR-Tools for combinatorial optimization with a custom gradient-boosted model trained on 6 months of GreenRoute historical delivery data. The model predicts actual travel times between locations based on time of day, day of week, weather, and seasonal patterns — far more accurate than Google Maps estimates for delivery vehicles in GCC traffic.
We also built a constraint engine that handles the unique GCC requirements: time windows for gated communities, cold-chain routing that minimizes transit time for temperature-sensitive goods, driver skill matching (some deliveries require heavy lifting or special handling), and prayer time buffers built into driver schedules.
The driver app: designed for one-handed use
The Flutter driver app was designed with a single constraint: it must be fully operable with one hand while the driver is at a delivery stop. Large touch targets, swipe gestures for status updates, voice-activated navigation, and automatic delivery confirmation via geofencing. Drivers spend less than 5 seconds interacting with the app per delivery.
The app also provides real-time route adjustments. If a delivery fails (customer not home), the driver taps one button and the system instantly recalculates the remaining route, potentially rearranging stops to minimize backtracking. If traffic conditions change mid-route, the app suggests alternative paths without requiring dispatcher intervention.
Dispatcher dashboard: from 3 hours to 15 minutes
The Next.js dispatcher dashboard replaced the morning route planning ritual. Dispatchers now upload the daily delivery manifest (CSV or API integration with the client e-commerce platform), review the AI-generated routes on an interactive map, make any manual adjustments (VIP customers, special instructions), and push routes to all drivers with a single click.
The dashboard also provides real-time fleet visibility: driver locations, delivery progress, predicted completion times, and automatic alerts for delays. This real-time view replaced a WhatsApp group where dispatchers previously tracked drivers manually.
Results
After 3 months in production, the numbers speak clearly: average delivery time reduced by 28% (from 45 minutes between stops to 32 minutes), fuel costs down 19% across the fleet, on-time delivery rate improved from 82% to 96%, and morning route planning time reduced from 3 hours to 15 minutes.
Driver satisfaction also improved significantly. The optimized routes are more logical (less backtracking), the app reduces phone calls from dispatchers, and drivers finish their routes earlier. Driver retention improved by 22% in the first quarter — a major cost savings in an industry with notoriously high turnover.
GreenRoute has since expanded the system to their Riyadh and Doha operations and is exploring adding predictive demand modeling — using historical data to pre-position inventory at satellite locations before orders arrive.
Lessons learned
Historical data beats theoretical models. Our travel time predictions trained on actual delivery data outperformed generic routing APIs by a wide margin because they capture the reality of delivery driving — parking time, building access, customer interaction — not just road travel time.
Regional context matters enormously for logistics. Prayer time scheduling, Ramadan traffic patterns, extreme heat routing constraints — these are not edge cases in the GCC, they are core requirements. Any logistics optimization that ignores them will fail in production.