Harborline Insurance
A predictive risk-scoring model that flagged high-risk claims before payout, cutting fraud losses 18%.
Industry
Finance & Insurance
Timeline
12 weeks
Team
2 ML engineers, 1 data engineer
Category
AI Development
Where things stood
Harborline's fraud team was reviewing claims manually after payout, which meant confirmed fraud losses were already out the door by the time a pattern was caught.
How we got there
Historical data pipeline
Built a pipeline over five years of claims history to train on, cleaning up label inconsistencies that would have skewed the model otherwise.
Risk-scoring model
Trained a model to flag high-risk claims before payout rather than after, scoring claims as they enter the queue.
Human-in-the-loop review
Routed flagged claims to a human reviewer instead of auto-denying, since a false positive denying a legitimate claim is its own cost.
What changed
18%
Reduction in fraud losses
Pre-payout
Flagging now happens before payout, not after
<5%
False-positive rate on flagged claims
Tech stack
Python
The language for the data pipeline and model training.
PyTorch
Trains the risk-scoring model on five years of claims history.
scikit-learn
Handles preprocessing and the baseline models used to validate the deep model's lift.
AWS SageMaker
Runs training and serves real-time risk scores as claims enter the queue.
PostgreSQL
Stores claim records and model scores for the fraud team's review queue.
Docker
Packages the training and inference pipeline consistently across environments.
“The model catches patterns our adjusters were trained to look for, but faster and at real scale. It paid for itself in the first quarter.”
Anita Shah
VP of Claims, Harborline Insurance
