Discovery
- Ride-alongs with 12 drivers across 4 cities
- Mapped 9 dispatcher workflows
- Surfaced the real bottleneck: re-routing latency
Logistics · 2024
We taught a fleet to think.
The challenge
ShipFast was hand-routing 2M+ daily packages across 14 cities. Operations was the bottleneck. Drivers idled, deliveries slipped, customers churned.
The approach
We built a reinforcement-learning routing engine over a hybrid graph of road network, weather, and historical pickup data — running continuously, re-optimising every 90 seconds against live conditions.
By the numbers
$0
Saved year one
0%
Fewer miles driven
0%
On-time rate
0M
Daily routes solved
The process
From discovery to launch.
Architecture
- Graph DB for road network + delivery constraints
- Event-driven re-optimisation pipeline
- Edge inference for sub-second routing decisions
Design System
- Dispatcher UI rebuilt around live route stream
- Driver app with offline-first routing
- Component library for ops dashboards
Engineering
- PyTorch RL model with custom reward shaping
- Go-based routing service, p99 < 80ms
- Kafka for delivery state, Postgres for history
Launch
- Shadow-mode rollout city by city
- On-call rotation co-staffed for 60 days
- Migration playbook handed off to internal team
Driver app · dispatcher console
Their team understood our logistics domain immediately. The route engine they built saved us $4M in year one and the ROI was clear from week two. Best engineering partnership we've ever had.
Priya NairVP Engineering, ShipFast
4M
Cost saved
31%
Miles cut
99.94%
On-time
Outcomes
The numbers settled. The story didn't.
Within two quarters the routing engine had paid for itself three times over. Dispatcher headcount stayed flat as volume grew 38%. Driver retention rose from 71% to 89%.
Next case study