TerraFleet

Client / Project

TerraFleet is a 60-person SaaS company providing fleet-management software to mid-market logistics operators. Their customers run between 50 and 500 vehicles each. Every vehicle streams telemetry — engine temperature, oil pressure, fuel efficiency, GPS, hard-braking events — into TerraFleet's pipeline.

Problem

That data was sitting in their warehouse mostly unused. Customers were churning at renewal because competitors had started offering predictive maintenance and TerraFleet had no equivalent. The internal roadmap was full and the data-science org existed only as a slide; standing it up was a year away.

Approach

A 3-engineer Project Pod for 9 weeks. We chose a deliberately unfashionable architecture: not an ML model, but an agent that streams telemetry windows and reasons about them against a runbook of failure patterns we built from the mechanics on staff at three friendly customers. The agent flags conditions, ranks them by predicted urgency, and explains in plain English why a given vehicle should go to the shop this week instead of next month.

We chose this approach because it was shippable in 9 weeks without a training dataset, and because the explanations made it acceptable to customers in a way a black-box ML score never would have been. Fleet operators trust an explanation they can argue with.

Outcome

The system flagged 73% of pre-failure conditions an average of 4 days before vehicles actually broke down. Across the three pilot customers, vehicle-downtime hours dropped roughly 22% over the first quarter. The feature shipped to general availability at renewal season and was credited with retaining two accounts that had been openly evaluating competitors. TerraFleet still has not hired a data scientist.