Fraud Detection

01Where in the payment flow does this sit — inline, or after the fact?

Inline: the gateway calls risk scoring synchronously during authorization and waits for the verdict — a decision made after the money moved is an alert, not a defense.

02Is the decision binary — approve or block?

Three tiers: approve, step-up challenge (3DS or OTP) for the uncertain middle, and block only at high confidence — the challenge tier lets a real customer prove themselves instead of being declined.

03Pure ML, pure rules, or both?

Both in one call: the model scores every transaction, and a rules engine can hard-block, force-review, or allow-list on top — rules are the fast lever the model cannot be.

04What does the model actually see at decision time?

Precomputed velocity, device, and graph features fetched from a feature store, plus a few cheap request-time features (amount, merchant category, geo mismatch) computed on the spot.

05How does the system ever learn it was wrong?

A label pipeline: chargebacks and disputes arrive 30–90 days later, analyst verdicts arrive in hours, and both flow back — each decision also stores its exact feature snapshot so training sees what the model saw.

06Who looks at the borderline cases?

A case review queue: mid-risk and rule-flagged transactions go to human analysts with score, reason codes, and account history — their verdicts are the fastest labels the system gets.

01How many milliseconds do we own inside the authorization?

Under 100 ms of the ~2-second authorization belongs to risk; the scoring service budgets 50 ms p99 so one slow dependency cannot blow the slice.

02The feature store times out mid-transaction — what happens?

The payment is never held hostage: on a feature-store timeout the service degrades to rules plus a default score and flags the decision — fail-degraded, not fail-closed.

03How fresh do velocity features have to be?

Velocity features are seconds fresh: a stolen-card burst must show up in the counters by the third swipe, so aggregates come from a stream processor, not a nightly batch.

04What metric is this system actually judged on?

The system is judged in dollars, not accuracy: at a 10-basis-point fraud rate a model that approves everything is 99.9% accurate — the real trade is false-decline revenue against chargeback loss.

05Can we explain any single decline afterwards?

Every decision persists its score, top feature reasons, rule hits, and model version — analysts, disputes, and regulators all ask why, and the answer cannot be a bare score.