Centurion · Fraud · ML Graph
Real-time graph. Purpose-built for fraud and abuse at scale.
Score every transaction, login, and signup against the live fraud graph — in real time, at the moment of decision, not in tomorrow's case-review queue.
Millions
entities per query
< 1s
graph traversal
Real-time
scoring response
Online
not batch case review
Why real-time matters
Offline graph analytics catch fraud after the loss.
Traditional graph tooling — Neo4j, batch GraphX jobs, Spark notebooks — is built for case review. You query the graph hours or days after the event, surface a ring, and write it off. The money has already left.
ML Graph runs at decision time. When a transaction hits, when a signup completes, when a login is attempted — ML Graph scores the entity's place in the network in real time, returns a verdict, and the decisioning system approves, protects, steps up, or blocks. The loss never happens.
That latency profile is what unlocks graph as a preventative signal instead of a forensic one. It's what lets a survival model track a tainted asset through the network before the next chain-of-events transaction posts. It's what makes graph-based fraud detection a production system, not a quarterly investigation deliverable.
What's inside
Built for online decisioning.
Real-time ingestion
Transactions, signups, logins, payments — written into the graph as they happen. The graph state during the scoring call is the live state, not a stale snapshot.
Online neighborhood queries
Multi-hop ring detection, ego-network statistics, and path-existence checks run live at decision time — not as batch jobs after the fact.
Millisecond inference
GNN and traversal-based scoring at decisioning latency. The transaction-level scoring call returns before the user notices a pause.
Purpose-built for fraud and abuse
Coordinated rings, money mules, bust-outs, synthetic identity, promo abuse, mass-account creation. The fraud patterns are the design target — not a generic graph use case.
Survival models on the network
Track tainted assets, accounts, and devices over time. Watch the signal propagate through the network — and intervene before the next event.
Explainable enough for risk committees
Every score returns the supporting subgraph and feature attribution. The signal can show its work to an auditor or a regulator.
In production
ML Graph supports real-time detection of bot and ghost-broker attacks across online personal lending and P&C insurance, surfacing coordinated fraud at the moment of application.