Blog

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

Common Mistakes Adding an LLM Layer to a Legacy Fraud Engine

The recurring LLM fraud engine mistakes when adding a reasoning layer to legacy rules: no shadow mode, all-traffic calls, ungrounded output, no fallback path.

5 Fraud Detection Signals Beyond Transaction Amount

Transaction amount is a weak fraud signal alone. Five fraud detection signals that do more work - velocity, device, graph, behavior, text - and how each fails.

How Real-Time Fraud Scoring Systems Are Architected

Real-time fraud scoring is an architecture problem: the latency budget, the online feature store, graceful degradation, and the async audit path.

Rules Engine vs LLM Fraud Scoring: What Actually Changes

A rules engine, a gradient-boosted model, and an LLM layer each change something specific in fraud scoring. What actually shifts, and what doesn't.

What Counts as Explainable AI in Fraud Detection?

Explainable AI in fraud detection means more than SHAP values. What a fraud decision has to explain, to whom, and where post-hoc explanations fall short.

Build vs Buy Fraud Detection: When to Build In-House

Most teams should buy fraud detection first. The specific pressures - false-positive cost, proprietary signal, volume, compliance - that justify building.