Streaming ML platform scoring 2 million card transactions/minute with sub-100ms latency.
Real-time fraud platform — overview
Client
Top-10 US Retail Bank (confidential)
Location
New York, NY
Timeline
1 Feb 2023 — 15 Aug 2024
Industry
Financial Services
Streaming ML platform scoring 2 million card transactions/minute with sub-100ms latency.
Engineered a Kafka + Flink + Feast streaming feature platform delivering sub-100ms fraud-score decisioning at peak loads of 2M transactions/minute. Reduced false positives by 22% and lifted fraud catch rate by 14%.
Gallery
Execution timeline
Live progress tracking with photographic documentation from each construction phase.
Phase 01
Kafka topic design, schema registry, and exactly-once semantics across producers.
Phase 02
Real-time + offline parity feature store with Feast and online Redis serving.
Phase 03
Champion / challenger model rollout with shadow scoring and bias monitoring.
Phase 04
Sub-100ms SLA, autoscaling, chaos drills, and 24x7 SRE coverage.
Related work
Migrated 30+ legacy data warehouses into a unified lakehouse, cutting query latency 8x and reporting costs 40%.
Forecasting + assortment optimization platform increased gross margin by 3.1% across 1,400 stores.
Eliminated error-prone in-memory reporting by migrating to a Snowflake data warehouse — error turnaround dropped from days to hours, and end users noticed immediately.
Tell us your business question, your data estate, and the timeline. We respond within one business day with a written hypothesis and a delivery plan.
Enquiries
contactus@vitanagroup.com
Support
support@vitanagroup.com
