Product Context

FinFlow is a digital payments platform that must make a real-time decision on every card transaction: approve, decline, or send for step-up verification. The service is used by merchants and consumers globally, and model quality directly affects fraud loss, conversion, and customer trust.

Scale

Task

Design a low-latency, real-time ML decision service on AWS using Python and SageMaker. Your design should address:

1. How requests flow from API ingress through feature lookup, candidate decisioning, model scoring, policy checks, and response.
2. What should run online vs batch, including feature computation, model training, model deployment, and backfills.
3. Which models you would use at each stage, and why a multi-stage architecture is preferable to a single model.
4. How you would evaluate the system offline and online, including delayed labels for fraud outcomes.
5. How you would handle monitoring, feature drift, training-serving skew, and safe rollback.
6. What failure modes you expect at this scale and how the system should degrade gracefully.

Constraints

• The service must return a decision within 80ms p99; hard timeout at 100ms.
• False declines are expensive, but missed fraud has direct financial cost; the platform wants tunable thresholds by merchant risk tier.
• Some labels are delayed by days to weeks due to chargebacks, so training data is partially censored.
• PII and payment data must remain compliant with PCI requirements; minimize sensitive data movement.
• The system must continue serving during partial outages in feature infrastructure or model endpoints.
• Cost matters: the online path should avoid heavyweight per-request deep models unless justified by measurable lift.