An AI-native routing engine that cut delivery costs by ~18%.

The challenge

A regional logistics operator was routing delivery vehicles using a rules-based system that hadn't been meaningfully updated in years — dispatchers manually adjusted routes around traffic and driver availability, and fuel costs were climbing faster than delivery volume.

The approach

Rather than replace the entire dispatch system, we built a routing-optimization layer that sat alongside it — ingesting live traffic data, vehicle capacity, and delivery windows, then proposing optimized routes that dispatchers could accept or override. Keeping a human in the loop was a deliberate choice: dispatchers had context the model didn't, and trust in the system depended on it being adjustable, not a black box.

What we built

A Python-based optimization engine served through a FastAPI backend, integrated into the client's existing dispatch dashboard rather than replacing it. The model re-optimizes routes as conditions change through the day, and every recommendation includes the reasoning behind it — which stops were reordered and why — so dispatchers could sanity-check output rather than blindly accept it.

The result

Delivery costs dropped by roughly 18% within the first full quarter after rollout, driven mainly by reduced fuel usage and fewer late deliveries triggering compensation clauses. Dispatcher override rate on suggested routes dropped steadily over the first two months as trust in the system's recommendations grew.