Design Fraud Detection System - System Design | WinJob

Problem Statement & Context

You are designing a real-time fraud detection platform for card-not-present payments — the decision layer Stripe, PayPal, and Sift operate at authorization time. This section focuses on mission, stakeholders, and authorization-time constraints. Interviewers expect you to separate online scoring (must finish before auth timeout) from offline learning (labels, graph expansion, rule tuning).

1.1: score card-not-present auth in <100 ms p99
1.2: balance fraud loss vs false decline revenue
1.3: feed analyst queues without drowning SOC

Mechanism

Merchants send AuthorizationRequest events with PAN token, amount, device, billing/shipping signals. The platform returns allow | review | block plus reason codes. High-risk traffic routes to 3DS step-up; blocked traffic still logs features for chargeback defense.

Interview phrasing

Lead with numbers: 180K TPS peak, 85 ms scoring budget, <0.3% false-positive SLO, <2% traffic manual review ceiling. Mention PCI scope reduction via tokenization and that models never see raw PAN.

Failure mode to volunteer

During Black Friday, feature store lag hits 4 minutes; velocity rules misfire. You degrade to rules-only path, widen review band, and page feature platform—not payments.

Signal	Target
Peak TPS	180K TPS
Score p99	85 ms
False positive	<0.3%
Availability	99.95%

javaOne Dark Pro

1 public record AuthContext(String txnId, String cardToken, long amountCents, String deviceId) {}

pythonOne Dark Pro

1 def decision_bucket(score: float, review_hi: float = 0.55, block_hi: float = 0.82) -> str:
2     if score >= block_hi:
3         return "block"
4     if score >= review_hi:
5         return "review"
6     return "allow"

typescriptOne Dark Pro

1 export type FraudDecision = "allow" | "review" | "block";
2 export interface ScoreResult { score: number; decision: FraudDecision; reasonCodes: string[]; }

Why interviewers care

Fraud Detection System interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Problem Statement & Context that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Problem Statement & Context: mission, stakeholders, and authorization-time constraints
•Ops angle: feed analyst queues without drowning SOC
•Metric anchor: 180K TPS / 85 ms

1	def decision_bucket(score: float, review_hi: float = 0.55, block_hi: float = 0.82) -> str:
2	if score >= block_hi:
3	return "block"
4	if score >= review_hi:
5	return "review"
6	return "allow"

1	export type FraudDecision = "allow" \| "review" \| "block";
2	export interface ScoreResult { score: number; decision: FraudDecision; reasonCodes: string[]; }