AME Digital builds a machine learning fraud detection model on Databricks reaching 90% accuracy

Overview

AME Digital, a Brazilian fintech digital wallet with over 33 million customers, used Delta Lake, MLflow and Unity Catalog on the Databricks Data + AI Platform to unify over 700 previously siloed datasets and build a fraud detection algorithm that analyzes transactions in real time to flag suspicious behavior before it occurs. The model reached 90% accuracy in fraud prevention, ran 3.2x faster than the prior on-premises pipeline, and helped cut operational costs by 34% and job execution time from 5.5 hours to 50 minutes.

The challenge

AME Digital managed over 700 datasets often isolated in data silos within different teams and systems, on a legacy on-premises Hadoop environment that was complex, costly and of limited scalability, which caused error-prone, resource-intensive workflows, hindered predictive fraud solutions, created regulatory compliance challenges, and drove up storage and compute costs.

The solution

AME Digital chose the Databricks Data + AI Platform for its unified environment, using Delta Lake to centralize over 700 datasets, MLflow to manage machine learning cycles for a fraud-catching algorithm, Unity Catalog to manage and secure PII and access controls, Photon for large-scale query performance, and Power BI integration for reporting, with Databricks-certified partner Eleflow Big Data helping analyze data in real time and migrate from 1TB to over 400TB of data.

Reported business value

AME Digital's fraud detection model reached 90% accuracy, job execution times were cut from 5 hours 30 minutes to 50 minutes (an 85% time reduction), operational costs fell 34%, and the model pipeline ran 3.2x faster than on-premises.