Top-10 Bank Saves $28M/Year with Real-Time Fraud Detection on Databricks
Replaced a brittle rules-based fraud system with a real-time ML platform on Databricks + Snowpark, scoring 50M+ transactions daily.
A top-10 US bank was running an aging rules-based fraud engine that produced thousands of false positives a day — degrading customer experience and missing increasingly sophisticated fraud patterns. Each new typology took the risk team 6+ weeks to model, deploy, and validate, by which time the attackers had already moved on.
Apptad built a real-time ML fraud detection platform on Databricks with feature pipelines feeding from Snowflake, Snowpark-based scoring, and MLflow-tracked model lifecycle. We co-designed the feature store with the bank's risk team, set up automated retraining triggered by drift, and integrated with the core authorization stream for sub-100ms scoring on every card-not-present transaction.
The new platform scores 50M+ transactions per day with sub-100ms latency and a 94% true-positive rate. False positives dropped 71%, time-to-deploy a new model fell from 6 weeks to under 4 days, and the bank attributes $28M in annual avoided fraud loss to the system in its first full year.
The Numbers Behind the Story
Built On
The platforms and partners that powered this engagement. Where Apptad has a dedicated alliance, the chip links to our partnership page.
“Apptad's data team didn't just deliver a model — they delivered the full operating model around it. Fraud is now an engineering problem we solve, not a tax we pay.”
