// the take
An AWS-published reference architecture showing how to detect credit card fraud with Amazon SageMaker, combining a supervised XGBoost classifier with an unsupervised RandomCutForest anomaly detector, then exposing both behind a Lambda/API Gateway endpoint. It's aimed at AWS solution architects and ML engineers who want a worked example of wiring SageMaker training and inference into a real request pipeline, not at people looking for a fraud-detection library to drop into their stack.
Runs supervised and unsupervised models side by side so it covers both the labeled-data case and the common fraud scenario where you only have unlabeled or delayed-label transactions. Explicitly handles class imbalance with imbalanced-learn upsampling and XGBoost's scale_pos_weight rather than ignoring it, which is the detail that actually matters for fraud data. The inference path is a full shape, not just a notebook: API Gateway and Lambda front the SageMaker endpoints, and predictions get written back to S3 through Kinesis Firehose for monitoring or retraining.
The README tells you outright not to clone and run this manually — it's built to be launched through SageMaker JumpStart's 1-click flow, a distribution mechanism AWS has been winding down, so the supported path may not stay available. Default sizing in the cost table runs about $900/month (ml.c5.12xlarge Studio instances, always-on endpoints), which is steep for something meant to be a learning demo. The dataset is the well-worn 2013 ULB credit card set with PCA-anonymized features, so there's no real feature engineering for velocity, device, or graph signals that actual fraud systems rely on — this teaches the SageMaker plumbing more than fraud detection itself. No commits since October 2024 and it's maintained as an AWS sample, not a project with ongoing support.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →