Transaction Scoring

Score a single transaction against the PyTorch fraud classifier and autoencoder anomaly model. High or uncertain results open a review case.

Download demo CSV
Review pipeline
1
Score
PyTorch classifier + autoencoder produce risk and anomaly signals.
2
Route
Low → audit record only. Uncertain / High → review case opened.
3
Prepare
LangGraph agent retrieves policy context, generates grounded brief.
4
Decide
Human analyst records rationale and action. Audit trail persists.
Risk bands
high Above high-risk threshold or severe anomaly — review case opened
uncertain Above review threshold — review case opened
low Below review thresholds — audit record only
Field reference
Time — seconds since dataset start
Amount — transaction amount
V1–V28 — anonymized PCA features, not business-readable
High scores mean similar to fraud patterns in training data — not confirmed fraud.