How Persistent proposed a Dataiku and Neo4j-led fraud intelligence platform to improve real-time client identification, device resolution and fraud decisioning.
~50%
Current accuracy baseline
60%
Pilot accuracy target
Up to 75%
Scale-out accuracy target
>250 ms
Inference latency target
The client was looking to strengthen fraud detection across client identification and device resolution journeys where decisions had to be made in real time. Its existing vendor-based approach delivered accuracy of around 50%, creating avoidable false positives and leaving some fraudulent activity undetected.
The organization had already built a fraud data mart on the SATP platform and onboarded most of the data sources needed for client identification and device resolution. The next step was to turn that foundation into an intelligent, governed and low-latency fraud intelligence capability powered by AI, machine learning and graph-based identity resolution.
Persistent proposed a hybrid architecture built on Dataiku and Neo4j. Dataiku would provide the AI and MLOps foundation for feature engineering, model training, validation, deployment, monitoring and real-time inference. Neo4j would add the relationship intelligence needed to connect identities, accounts, devices, IP addresses, sessions and transactions into a more complete fraud-risk view.
When Static Fraud Scoring Could No Longer Keep Pace
Fraud detection in digital financial services is no longer a rules-only challenge. Fraudsters rarely create risk through isolated events. They move across devices, locations, beneficiaries, sessions and transaction patterns, making it difficult for static rules or point-in-time scores to identify connected risk with enough precision.
Client’s existing fraud detection environment was constrained by accuracy and context. The solution delivered only around 50% accuracy, which meant the business had to manage both false positives and missed fraudulent activity. For teams responsible for live client identification and device resolution, that created a direct operational challenge: detect more fraud without slowing down legitimate customers.
The cost of inaction was clear. Without a connected intelligence layer, fraud teams would continue to depend on isolated model scores, fragmented signals and manual investigation effort. That increased the risk of undetected account takeover, fraudulent transactions, social engineering-induced activity, unusual trading behavior, bank kiting and new device or location anomalies.
The client needed a fraud intelligence ecosystem that could combine real-time ingestion, historical pattern analysis, supervised models, anomaly detection, graph-based identity resolution, explainability, drift monitoring and latency engineering. The business requirement was demanding: improve on the 50% baseline while keeping inference latency below 250 milliseconds.
A Dataiku and Neo4j Model for Connected Fraud Intelligence
Persistent shifted the fraud detection approach from device risk scoring to a multi-layer fraud intelligence model. The proposed architecture kept deterministic rules where they added value, strengthened predictive risk scoring through machine learning and used graph intelligence to uncover hidden relationships that traditional models could miss.
Dataiku as the AI and MLOps foundation: Dataiku would support data ingestion, feature engineering, SQL-based transformations, model training, hyperparameter tuning, validation, deployment, API endpoints and monitoring. Its model registry and MLOps capabilities would help create a governed lifecycle from experimentation through production scoring.
Neo4j for relationship-aware fraud detection: Neo4j would provide the graph layer needed to model relationships across devices, accounts, identities, IP addresses and behaviors. These graph-derived features would help detect account takeover, synthetic identity patterns, mule networks, shared-device fraud and hidden linkages across customer activity.
Real-time inference with responsible controls: The platform was designed to support real-time scoring while maintaining governance through validation packages, explainability, performance dashboards, Responsible AI guardrails, drift monitoring, CI/CD pipelines, model registry and versioning.
A full fraud detection lifecycle: Together, Dataiku and Neo4j would support data ingestion, feature creation, model training, backtesting, real-time scoring, serving and drift monitoring. Each stage would include checks for data quality, feature completeness, leakage prevention, label quality, out-of-time testing, champion-challenger validation, score calibration and drift monitoring.
Fraud Intelligence Built for Real-Time Decisioning
The documented business trajectory is clear: move from an accuracy baseline of around 50% to a pilot target of 60%, with scale-out ambition of up to 75% accuracy subject to discovery validation. The platform also targeted inference latency below 250 milliseconds, allowing fraud decisions to operate inside live client identification and device resolution journeys.
With Persistent, the client would be positioned to improve fraud detection effectiveness while reducing unnecessary customer friction. Higher model accuracy would help fraud teams identify stronger risk signals. Graph intelligence would give investigators a more connected view of identities, devices, accounts and transactions. Dataiku-led MLOps would create the governance structure needed to monitor accuracy, drift and model performance over time.
The proposed platform would also improve fraud investigation by exposing relationship-based risk patterns. Instead of reviewing individual alerts in isolation, fraud teams could analyze how devices, IPs, accounts, beneficiaries and transactions relate to one another. This would make the fraud operating model more intelligence-led, explainable and scalable.
With Persistent, the client would be able to:
- Increase fraud model accuracy from an approximately 50% baseline to a 60% pilot target, with a scale-out ambition of up to 75% accuracy subject to discovery validation.
- Maintain sub-250 millisecond inference latency to support real-time fraud decisioning during client identification and device resolution.
- Use Neo4j relationship intelligence to improve visibility into account takeover, synthetic identity patterns, coordinated fraud networks and shared infrastructure.
- Create a governed MLOps foundation through Dataiku for model development, deployment, monitoring, registry, explainability and versioning.
- Reduce reliance on static rules and isolated model scores by combining deterministic controls, machine learning, anomaly detection and graph-derived features.
From Fraud Detection to Connected Fraud Intelligence
Persistent helped define a future-ready fraud intelligence platform that moves the client beyond static device scoring and into a connected model of customer behavior, device context, identity relationships and transaction patterns.
By combining Dataiku’s AI and MLOps lifecycle capabilities with Neo4j’s graph intelligence, the proposed platform gives fraud teams a stronger foundation for real-time decisioning, explainable risk detection and governed model operations. For the client, the transformation is not only about improving model accuracy. It is about building a fraud intelligence capability that can scale with the speed, complexity and trust expectations of digital financial services.
