LogisticsException handlingFreight and logistics operator
Freight document exceptions triaged by type, severity and next action
type, severity and action predicted together
- Status
- In production
- Built with
- PyTorchMulti-task learningSynthetic data
01The problem
A manual decision tree for document exceptions: a person reading each one, deciding its type, judging its severity and choosing the action, hundreds of times over.
02What we built
A multi-output model with a shared base and separate heads that predict exception type, severity and action together. No training set existed, so one was built from the documented decision process.
03The result
The manual decision tree became a model that handles routine cases and routes the ambiguous ones to a person. Building the training data was most of the work, which is usually true and rarely budgeted for.
More work
Something like this in your business?
Tell us how the business runs and where the time goes. We’ll tell you honestly what’s worth building, and what isn’t.