Blog

AI agents, RAG, and integrations — notes from production.

A Checklist for Choosing a Voice AI Vendor for Financial Services

A voice AI vendor checklist: measured phone-call latency, native barge-in, actions under your rules, a real warm handoff, and consent handling.

False Positives Are a Revenue Problem: Costing a Wrong Decline

The fraud false positive cost is invisible in the loss ledger: a wrong decline produces no chargeback, just lost margin and a customer who leaves.

Chargeback Fraud vs Account Takeover: Different Signals, Different Models

Chargeback fraud vs account takeover need different handling: one is a real-time authorization problem, the other a dispute-time evidence problem.

Card-Testing Attacks: How Velocity Features Catch a Burst of Small Charges

Card-testing attack detection is a pattern problem: each small charge looks fine, so the signal is velocity - attempts per card, device, and IP over minutes.

Real-Time Risk Scoring for Payments vs Lending: What Changes

Payments vs lending fraud scoring shares an architecture but not the constraints: sub-200ms authorization vs a regulated adverse-action decision.

The Fraud Analyst Console: Turning Human Overrides Into Training Data

A fraud analyst review console works the soft-hold queue, shows both model signals side by side, and logs every verdict as a training label.

Graceful Degradation: What a Fraud Model Does When the LLM Times Out

Graceful degradation in a fraud model: a named fallback for each failure - LLM timeout, slow feature store, model down - each one flagged, counted, and tested.

Why Kafka for a Fraud Pipeline — and When a Lighter Queue Is Enough

Kafka in a fraud pipeline earns its keep for one reason: a durable, replayable log to reconstruct what the system knew about a transaction months later.

Idempotency in a Real-Time Fraud Pipeline: Why Retries Don't Double-Score

Idempotency in fraud scoring: a transaction event can arrive twice, and without a dedupe key it inflates velocity features and duplicates the audit trail.

Point-in-Time Correctness: Building a Fraud Training Set That Matches Production

Point-in-time correctness means each training row uses feature values as of that moment, not as of now - the fix for the leakage that makes offline metrics lie.