Case study · Financial services · AI/ML and streaming

Real-time point-of-sale fraud detection for a leading credit card firm

A leading credit card firm moved point-of-sale fraud checks from slow, after-the-fact methods to a streaming platform that scores each transaction within 100 milliseconds.

Industry
Financial services, credit cards
Capability
AI/ML and streaming
Platform
Kafka, APIs, GPU clusters
Implementation period
8 months

Context

The client is a leading credit card firm. It asked VerticalServe for a fraud detection solution that could process and analyse large volumes of point-of-sale transaction data quickly, respond to potential fraud as it happens, and keep false positives low.

The problem

The firm could not reliably detect and prevent fraudulent point-of-sale transactions in real time. Its traditional fraud detection methods were slow and less effective, which led to significant financial losses and damage to the firm’s reputation.

What was built

VerticalServe built a detection path that scores each transaction as it happens, and a model lifecycle around it that keeps the models current.

Flow: account, bureau, merchant, time and location data stream through Kafka to a fraud detection API that scores within 100 ms; models are trained on GPUs, managed by MLOps pipelines with challenger models, and reviewed by fraud analysts. DATA SOURCES Account and bureau data Merchant details Time and geolocation REAL-TIME SCORING Kafka streaming Fraud detection API within 100 ms MODEL LIFECYCLE GPU training MLOps and challenger models Fraud analyst review Model monitoring · analyst feedback into the models

Flow: account, bureau, merchant, time and location data stream through Kafka to a fraud detection API that scores within 100 ms; models are trained on GPUs, managed by MLOps pipelines with challenger models, and reviewed by fraud analysts.

  • Data source integration: customer account data, credit bureau information, merchant details, transaction time and geolocation are combined into one view of each transaction.
  • Real-time detection through APIs: each transaction is processed and analysed within 100 milliseconds, with instant feedback on potential fraud.
  • Kafka for data streaming: high-throughput, low-latency streaming of large volumes of transaction data.
  • GPU clusters for model training: fraud models are trained on large transaction and authorization datasets.
  • MLOps pipelines and challenger models: training, deployment and monitoring are automated, and challenger models are evaluated continuously against the models in use.
  • Model monitoring and analyst review: an interface where fraud analysts review and validate model outputs, give feedback and adjust model parameters.

Stack

APIs Kafka GPU clusters MLOps pipelines Challenger models Model monitoring

How it was delivered

The engagement ran for eight months. The work covered two things: the real-time path (data source integration, Kafka streaming and the detection APIs) and the model lifecycle (GPU training, MLOps pipelines, challenger evaluation and monitoring).

Fraud analysts were part of the design. They review and validate what the models produce through a purpose-built interface, and their feedback goes back into the models.

The architecture is cloud-based and modular, so it can scale and be adapted as business needs and fraud patterns change.

The result

ms
Detection and response time, down from minutes
90%
Reduction in false positives
70%
Reduction in losses from credit card fraud
  • Time to detect and respond to potential fraud dropped from minutes to milliseconds.
  • False positives fell by 90%, which reduced disruption for genuine customers.
  • Financial losses from credit card fraud fell by 70%, saving the firm millions of dollars a year.
  • MLOps pipelines and the working loop with fraud analysts keep the models improving as fraud tactics change.
  • The firm has since extended the solution to other payment channels and fraud types.

The figures above are as reported in the original write-up.

What it means for a similar team

  • Bring every signal about a transaction together before scoring it. The write-up credits the accuracy gain to combining account, bureau, merchant, time and location data with the technology used.
  • Design the model lifecycle with the scoring path, not after it. Challenger models and monitoring are what kept detection effective here.
  • Give fraud analysts a way to review and correct the models. Their feedback was part of the system from the start.

This case study was first published on our Medium publication. Original article

Engage

Start a similar POD

Tell us the outcome you want and your constraints. We will reply with a discovery proposal.

Talk to us

Working on something similar?

An Innovation POD is a dedicated team that builds inside your environment. Tell us what you need to ship.

Start a POD