The workflow before#
This work spanned three payment companies with more than 100 million customers between them. Cash App, part of Block, had data spread across Redshift, EMR, Glue and S3, and folding in data from Afterpay, which Block acquired, added another layer. Fraud signals and business reports ran hours behind. At PicPay, which serves 60 million customers in Brazil, overnight analytics jobs ran into the next business day, data scientists spent about 70% of their time maintaining infrastructure, and the cash-back program cost far more than it needed to. In the Philippines, Coins.ph waited weeks for each round of ML experiments, so it was slow to respond to new fraud patterns. All three faced tighter regulatory scrutiny at the same time.
What we built#
We built each company a lakehouse on Databricks that brings scattered data into one governed place, with fraud models and analytics running on the same data. The core design was shared across all three: reliable storage with built-in quality checks, one catalog that controls who sees what, and a standard path for building, deploying and retraining models. We moved in phases so payments never stopped.
Quality checks before data moves on
Delta Lake gives each company transaction-safe storage, which matters when the data is money. Automated checks validate more than 200 business rules as data lands and catch anomalies before they reach a report or a model. Analysts stop discovering problems in a dashboard after an executive has already seen it.
One catalog for access and compliance
Unity Catalog sets access, lineage and audit trails in one place, with role-based permissions for more than 500 users across data science, analytics and engineering. Metadata tags detect personal data automatically and apply encryption. We put governance in the catalog rather than in each pipeline, so a new dataset picks up the rules the moment it's registered and project approvals no longer wait weeks on manual sign-off.
Fraud models that score in milliseconds
We migrated more than 50 existing models onto MLflow and set up automated retraining, which cut model drift by 60%. Fraud models read more than 200 features and score each transaction in under 100 milliseconds, across more than 10TB of transaction data a day. Detection accuracy reached 90%, and data scientists now run an experiment in hours instead of weeks.
Self-serve analytics for the business
The last phase put governed self-serve tools and more than 50 dashboards in front of business users. Self-serve users grew 25%, and fewer routine requests landed on the data teams. We built the dashboards on the same governed tables the models use, so the business and the fraud team never argue over whose numbers are right.
How we delivered it
We ran the work in four phases over six months, starting with fraud detection because it carried the most risk and the fastest payback. Phase two unified customer data, phase three tuned the analytics pipelines and phase four opened self-serve access. There was no single big cutover, and the first results showed up inside 30 days. Weekly training sessions and a center of excellence at each company meant the internal teams could run and extend the work themselves.
Before and after#
| Measure | Before | After |
|---|---|---|
| Core data job runtime | 5.5 hours | 50 minutes |
| ML experiment cycle | Weeks | Hours |
| Analytics infrastructure cost | Baseline | 50% lower |
| PicPay cash-back program cost | Baseline | $10M a year saved |
| Fraud detection accuracy | Fragmented signals across systems | 90% |
What's next#
With governed data and a standard model path in place, each company can push further into personalization and cross-sell. PicPay's sharper segmentation has already lifted customer lifetime value 15%, and Cash App can now put Afterpay data to work in cross-sell. New models reuse the same features and controls without new infrastructure.
