Real-time routing for a 1,200-vehicle fleet
A streaming pipeline and optimization service that replans routes continuously instead of once each morning.
- Lower fuel cost per drop
- 18%
- Events processed per hour
- 1.2M
- p99 replan latency
- 340ms
The challenge
Routes were planned overnight in a batch job and never revisited. A single traffic incident or late pickup invalidated a driver's whole day, and dispatchers spent their mornings manually rerouting over the phone. The batch job had also grown to seven hours, which put it at risk of not finishing before the shift started.
Our approach
Introduced a streaming ingestion layer for telematics and order events so the system had a live picture of the fleet rather than a nightly snapshot.
Extracted routing into a standalone service with a defined optimization objective, which made it possible to test route quality against historical data before going live.
Shipped in shadow mode for six weeks — the new routes were computed and scored but not issued — so we could prove the improvement before any driver was affected.
Gave dispatchers an override path, since the model does not know about the things they know.
The outcome
Replanning now happens continuously with a p99 latency of 340ms. Fuel cost per drop fell 18% and dispatcher intervention dropped by roughly two thirds. The overnight batch job was retired entirely.
Working on something similar?
If this looks like your situation, we can probably tell you within one call whether the same approach applies.