Fraud Detection

Fraud detection is a machine learning task that identifies suspicious transactions, account activities, or user behaviors in real time. Operators running local AI can deploy small, quantized models (e.g., XGBoost or lightweight neural nets) to score transactions on-device, avoiding cloud latency and data privacy risks. The model outputs a fraud probability (e.g., 0.95) or a binary flag; thresholds are tuned to balance false positives and false negatives. Local inference matters because fraud detection often requires sub-100ms response times and handling sensitive financial data without sending it to external APIs.

A local fraud detection pipeline might use a 50 MB XGBoost model quantized to 8-bit integers, running on an RTX 3060 with 12 GB VRAM. The model processes 10,000 transactions per second at ~0.5 ms each, flagging those with a score above 0.9 for manual review. This setup fits entirely in VRAM and avoids the 200 ms round-trip of a cloud API call.

In a typical workflow, an operator trains a fraud model using XGBoost or LightGBM on historical transaction data, then converts it to ONNX and runs inference with ONNX Runtime. The model is loaded into a Python script that reads streaming transactions from a Kafka topic, scores each one, and writes flagged IDs to a database. Operators tune the threshold by adjusting a config file and re-evaluating precision/recall on a held-out test set.