Catching Fraud
Before It Clears.
The rules missed $14M a year of coordinated attacks.
Each transaction looked clean; the graph of accounts
behind them didn’t. We built a temporal GNN (TGN / TGAT)
on Kafka Streams + Memgraph + Kùzu, with GNNExplainer
for analyst trust. 2.3M transactions a day, sub-800 ms,
99.6% precision.
Real-time fraud detection engine
A Series D fintech processing $8.2B a year had a fraud problem no rule set could catch. Each transaction looked fine; the rings behind them were bleeding $47M a year in patterns only visible on the graph.
We put a temporal graph neural network (TGN / TGAT on PyTorch Geometric + DGL 2.x) behind a Kafka Streams pipeline. It pulls 2.3M transactions a day, keeps the account-transfer graph live in Memgraph for streaming inference and Kùzu for analyst queries, and scores each transaction in under 800 ms via k-hop GraphSAGE-style sampling. Precision holds at 99.6% with 94% recall, and GNNExplainer / PGM-Explainer surface the offending sub-graph to the risk team.
Technical Approach
-
Transactions as a live temporal graph — Accounts are nodes, transfers are edges, and a temporal GNN (TGN / TGAT) on PyTorch Geometric + DGL 2.x learns embeddings that respect time. Rings that look clean one transaction at a time light up in embedding space — and Memgraph keeps the graph fresh enough to score on the live edge.
-
Streaming end to end — Kafka Streams on AWS MSK handles 2.3M transactions a day, and Flink CEP catches velocity spikes, geographic anomalies and ring-like chains before the money settles.
-
Precision before recall — A false positive blocks a real customer, and they don’t come back. We tuned the decision boundary to 99.6% precision first, then ensembled our way up to 94% recall.
-
Shadow, then cut over, then explain — We ran the GNN in shadow against the existing rules for 4 weeks. Once the score disagreements consistently favoured the GNN, we cut over with zero downtime — and wired GNNExplainer / PGM-Explainer into the analyst console so every block came with a sub-graph the risk team could read.
What was actually hard
A $200 transfer from A to B is unremarkable. A, B, C, D and E transferring to each other in a loop inside 90 minutes is a ring. The rules only saw one transaction at a time and the rings walked straight through them. We needed graph-level intelligence at sub-second latency — scoring every transaction before settlement without introducing a noticeable delay at the checkout.
Project Outcome
The GNN caught $12M of ring fraud the rules had missed entirely in the first quarter. False positives dropped 62%, so fewer real customers got blocked at checkout, and the risk team’s daily alert queue went from 3,000 down to 400 — and every one of them was worth looking at.
scoring > 99.6% precision at
94% recall > $47M annual fraud
prevented > 2.3M transactions
scored daily
“The GNN approach caught fraud patterns we didn't even know existed. Ring detection alone saved us $12M in the first quarter.”
@ Sarah K.
Head of Risk — Series D Fintech
More Case Studies
