Score a payment against everything the graph already knows

The same amount is routine from one account and alarming from another. Fourteen features are computed live from a NetworkX transaction graph — velocity, first-time counterparty, reverse edge, shared counterparties, account age — then a hand-weighted score and a small PyTorch MLP each get a say.

Feature, scoring and explanation code from graph-fraud-command-center · CPU inference · no GPU anywhere in this demo.

Synthetic throughout. The graph is generated by the repository's own simulator, and the MLP was trained on labels produced by a rule over that same synthetic stream — so its held-out accuracy measures "did the network learn the rule", not fraud-detection performance. Nothing here involves real financial data or is comparable to a benchmark.

Pick a payment

Sender

Receiver

Amount (USD)

Channel

From

The graph is live state: scoring a payment adds it, so pressing the button twice on the same pair flips is_new_pair from 1 to 0 and the score falls. That is the demo working, not drifting — history is exactly what these features are made of. The store rebuilds itself from a fixed seed every 150 scored payments, so the examples above always behave the same way for the next person to arrive.