TopN AnalyticsFraud Detection for Undisclosed Fleet Trips
A large EV fleet operator running electric vehicles under contract with a major ride-hailing platform, whose only visibility into vehicle usage came from that platform's own trip records — records that, by design, can't show trips run outside the registered platform.
Where things stood
Drivers are contractually required to use their assigned EVs exclusively for trips on one registered ride-hailing platform, but the operator's only visibility into usage came from that platform's own trip records.
Any usage outside the registered platform was functionally invisible, creating both a revenue leakage risk and an accountability gap.
The core difficulty: detecting something that, by design, has no record in the system being monitored — the operator needed an independent signal of actual vehicle usage that didn't depend on the ride-hailing platform's own data.
What we built
Vehicle Telematics (BigQuery), Ride-Hailing Trip Log (CSV), and Driver Check-in/out Logs (PostgreSQL) fed a Python/DuckDB join and feature-engineering layer.
A logistic regression model (scikit-learn, SMOTE-balanced) predicted whether a vehicle was 'on a trip' using only its own telematics — speed and the rate of change in battery state-of-charge — with no dependency on the ride-hailing platform's records.
Ground-truth labels were built via timestamp-window joins between telematics and trip records; logistic regression was selected over Random Forest and XGBoost alternatives for interpretability, with a decision threshold tuned to favor recall.
Predicted vs. platform-confirmed kilometers were reconciled per vehicle and driver into reporting outputs.
How it works
TopN Analytics built a supervised machine learning pipeline that compares what a vehicle's own telematics says happened against what the ride-hailing platform recorded — where telematics indicates driving activity but no matching trip exists in the platform's records, that gap becomes a flag for investigation.
The pipeline integrates three distinct data sources: onboard vehicle telematics, the ride-hailing platform's trip log, and driver check-in/check-out records that attribute a given vehicle session to a specific driver.
What we delivered and learned
The model achieved 77% overall accuracy and 82-83% recall on the 'trip' class on a held-out test set of over 400,000 telematics readings.
The mismatch was highly uneven — a subset of vehicles and drivers showed rates approaching half of all measured driving activity, concentrating investigative priority on a manageable subset.
TopN Analytics delivered concrete next steps: prioritize >20% mismatch cases, move to daily/live processing, and build a mismatch-tracking dashboard. The 23% figure is a model-based estimate of the scale of the problem, not a measured result of an intervention.
What it's built on
Want the full picture?
We're happy to walk through the details, numbers, and trade-offs directly.