FintechFintech & Payments

Real-Time Fraud Detection

A streaming detection model that flags anomalous transactions before settlement.

Client

PayShield

Location

Singapore

Engagement

12 weeks

Real-Time Fraud Detection — JSXDOM case study artwork for PayShield

<1s

flag latency

The story

From problem to proof.

PayShield processes millions of transactions daily across Southeast Asia. Their rules-based fraud system was losing an arms race: fraudsters adapted in weeks, rule updates took months, and false positives were blocking legitimate customers by the thousand.

We replaced the rigid rulebook with a streaming ML detection system that scores every transaction in under a second — learning new fraud patterns as they emerge.

What we brought

Streaming MLAnomaly DetectionPlatform Engineering

The challenge

Catch fraud in flight, not after settlement

Detection after the fact is accounting, not prevention. Scoring had to happen inside the transaction path — sub-second, at massive throughput — without blocking the 99% of customers who are legitimate.

  • Sub-second scoring inside the live transaction path
  • Fraud patterns that evolve weekly
  • False positives directly cost merchant revenue and trust

What we did

Four moves that made it work.

Streaming feature pipeline

Real-time features — velocity, geography, device fingerprints, behavioural history — computed in-flight for every transaction.

Ensemble detection

Gradient-boosted models for known patterns, anomaly detection for novel attacks, and the legacy rules demoted to a safety net.

Feedback flywheel

Analyst decisions on flagged transactions feed straight back into training — the system gets sharper with every review.

Explainable flags

Every flagged transaction arrives with the 'why' attached, cutting analyst review time from minutes to seconds.

The product

See what we shipped.

PayShield fraud console — live transaction stream with anomaly alerts
Live transaction stream with anomalies caught in flight
PayShield risk monitoring — model precision and false-positive trends
Model health: precision up, false positives down, every week

What we achieved

The numbers that followed.

Fraud losses dropped within the first month and kept falling as the feedback loop tightened. Legitimate customers stopped getting blocked — and support noticed before the dashboards did.

<1s

Flag latency in the transaction path

-63%

Fraud losses year over year

0.4%

False positive rate

9.2M

Transactions scored daily

“Our old system fought last year's fraud. This one learns next week's. The false-positive drop alone paid for the engagement.”

Raj Menon

VP of Risk, PayShield

Someone's case study like this is being written right now. It could be yours.

Every engagement above started with one free, no-obligation conversation.