Package-at-risk and loss-intensity models at FedEx
Senior Data Scientist · FedEx
Production ML across the logistics network. Identified at-risk shipments before they failed; quantified customer-level loss exposure.
Three years on the data-science team. The two models I’d point to:
Package-at-risk (88% accuracy)
Per-shipment classifier that scored the probability a package would miss its delivery commitment, given the customer’s historical shipping profile and the package’s current trajectory through the network. Ran in production on Azure with real-time inference, feeding operations dashboards.
Loss-intensity (90% accuracy)
Customer-level model — given a customer’s shipping mix and history, what’s our expected loss exposure over the next N days. Used by relationship managers and ops planning. Fewer inferences, longer-horizon, higher-stakes.
What was hard
The realtime constraint forced architectural choices the offline accuracy numbers don’t reflect. The first version of package-at-risk was a stacked ensemble that wouldn’t fit the latency budget; the production version traded a couple points of accuracy for a model that hit SLA every time. The right tradeoff, but only obvious in hindsight.
Earlier on the team I’d also done a GDPR-compliant entity-resolution model for customer records — 95% retrieval at the dedup threshold — and a fraud-detection model on shipping patterns that went into the legal team’s tooling.