Real-time detection at 50M daily transactions

Challenge:

20% annual churn killing growth targets. Unable to identify at-risk customers until cancellation notice received. Retention team reactive instead of proactive.

Solution:

Customer health scoring system analyzing 50+ signals: usage patterns, support interactions, feature adoption, engagement metrics. Identified churn risk 60-90 days before cancellation.

Results:

• 89% reduction in fraud losses ($10.7M saved annually)

• False positive rate: 78% → 12%

• Sub-100ms transaction processing.

• Scaled to 50M+ daily transactions (10x volume growth)

Technical approach:

• Neural networks + gradient trees ensemble

• Graph analytics for fraud ring detection

• Kafka for real-time event streaming

• TensorFlow Serving with Redis caching

Key takeaway:

Shadow mode deployment (monitoring without blocking) for 2 weeks validated performance before going live. Graph analysis exposed fraud rings missed by transaction-level models.