AI for Anti-Fraud, Anti-Money-Laundering and RegTech
Overview:
Across all the financial institutions, we see that modern economic crime becomes collusive,
cross-border, and professionally run — insiders working with external networks through layers
of companies and instruments — and rules built from expert memory cannot keep pace.
While financial crime has industrialized, the tools most financial institutions run today
were designed for an earlier generation of it, resulting in huge amount false positives and
missed detections.
Our approach consist of three ingredients, including: 1) semi-supervised machine learning,
which builds accurate models from few labels; 2) graph analytics, which makes relationships
the unit of analysis, and 3) human-in-the-loop design, leverages experts' judgment to refine
the models, keep evolving and defend against ever changing criminal activities.
We package this as four products covering the full AML workflow:
- Fraud and Money Laundering Network Detection on Graphs: Built graph-based
risk systems integrating customer, transaction, device, and geolocation data.
Applied semi-supervised learning, community and anomaly detection to uncover
coordinated fraud and laundering networks beyond rule-based coverage with
few labels; network visualizations let investigators trace collusion patterns.
Iterative refinement with expert feedback significantly improves detection of
organized, complex, ever-evolving and concealed financial crime.
- High-Risk Case Detection and Ranking: Engineered thousands of discriminative
features from customer profiles, transaction behavior, and historical cases
to build multidimensional financial-crime risk models. Developed a dual-engine
framework for recall expansion and intelligent ranking to surface high-risk
accounts missed by rules and prioritize existing alerts.
-
Predicate Typology Detection: Developed scenario-specific models for
underground banking, telecom fraud, online gambling, illegal fundraising, smuggling,
and other typologies. Built nearly 2,000 explainable features and 100+ rule-scoring models,
combining machine learning with expert feedback for risk scoring and typology attribution.
-
Automated Suspicious Activity Report Generation: Built AI-driven investigation
and reporting workflows using customer, transaction, and relationship data.
Applied NLP and LLMs to extract suspicious patterns, risk indicators, a
nd evidence and generate standardized reports, reducing manual effort and
omission risk while improving consistency and efficiency.
The technology and systems are widely deployed in
production with top financial institutions across mainland China.
They analyze over 200 million accounts every day, report more than one million suspicious
accounts, and helped identify and control over $20B in potentially illicit
funds every year — delivering over thirty percent more coverage at lower cost.
More detailed information can be found at
AHI Fintech, Inc