Your model works. It just hasn't seen the next variant yet.
FraudGen deploys adversarial AI agents to generate the synthetic fraud data your detection models are starving for.
Fraud detection has a data problem.
Few Fraud Transactions
Fraud is less than 0.1% of transactions. New variants are rarer still. No ML model can generalize from dozens of examples.
Mule Cases
A typical bank has 30 confirmed mule network cases. GNN training needs 1,000–5,000. The gap is an order of magnitude.
Unknown New Variants
Every new fraud variant starts with a detection gap — a window where the model has zero signal. The fraud continues undetected.
Our Solution
Blue Team
- Your existing fraud detection pipeline
- Retrains on synthetic data to close gaps
- Deploys hardened models with new coverage
Red Team
- AI agents that think like fraudsters
- Explore the full variant space at scale
- Run proactively before real attacks
Introducing the Admin Console
Admin observability and monitoring
See More Details →Mission Brief
Configure fraud type, variant count, and parameters
Coverage Matrix
Track fraud typology coverage and identify gaps
Fraud Constructor
Watch agents build transaction networks live
Quality Gate
Critic scores variants for realism and consistency
Operations Console
Monitor agent status, tokens, and throughput
Variant Journey
Trace each variant from seed through export
Real-Time Monitoring
Watch agents spawn and generate variants in real time. Persona cards stream in as criminal profiles are created. The coverage matrix fills live as the variant space is explored. Critic scores arrive with green/yellow/red quality indicators — and you can pause, resume, or stop the pipeline at any point.
Advanced Views
Heist Storyboard
Illustrated fraud scenario timelines
Quality Battlefield
Critic score comparison with drill-down
Pipeline Waterfall
Gantt-style pipeline execution timing
Built for Fraud Analysts
Architecture
From analyst input to ML-ready dataset — how the system works under the hood.
How It Works
A fraud analyst enters a natural language fraud description. The system decomposes it, generates diverse criminal personas, and deploys a parallel fleet of AI agents — each exploring a distinct corner of the fraud variant space. Every variant passes through deterministic validation (8-rule constraint checker, BFS graph labeling) before LLM-based quality scoring.
Agent Stack
Orchestrator
Coordinates the full generation pipeline end-to-end
Persona Generator
Creates diverse criminal profiles with realistic attributes
Coverage Matrix Manager
Tracks fraud typology coverage and identifies gaps
Fraud Network Constructor
Builds interconnected transaction networks
Schema Validator
Ensures generated data conforms to required formats
Critic Agent
Reviews and scores synthetic data for realism
Data Handler
Manages output formatting, batching, and export
Criminal Mastermind
Models adversarial decision-making with strategic multi-step planning
Mule Agents
Simulate money mule recruitment, layering, and cash-out behaviors
Bank System Agent
Emulates institutional controls, AML triggers, and detection responses
Level 4: Information Asymmetry — The Criminal Mastermind sees only what a real fraudster would see — no access to the bank's detection rules or thresholds. The Bank System Agent enforces real controls without knowing the criminal's strategy. This asymmetry produces realistic adversarial dynamics that statistical synthesis cannot replicate.
Data Pipeline
From description to dataset in minutes. Click any stage for details.
20–50
variants per run
≥ 7/10
mean critic score
~5–7 min
demo run time
~$4–6
per 20 variants