CLPS Incorporation Completes AI-Assisted Anti-Money Laundering Review Project for a Major Bank, Achieving Accuracy with Fine-Tuned Large Language Model
In response to the compliance challenges driven by continuously rising transaction volumes, the project innovatively explored the deep application of generative AI in AML review workflows. By fine-tuning small general-purpose Large Language Models (LLMs),
Significant Technical Milestones: Fine-Tuned Small General-Purpose Language Models Outperform Larger LLMs
In recent years, the complexity of transactions has surged, leaving AML review processes hindered by time-consuming manual workflows and highly subjective evaluation standards. Concurrently, the regulatory department has instituted stringent high requirements regarding the transparency, explainability, and evidence traceability of AI-assisted decision-making. Compounding these challenges, the Client's production environment operates with limited graphics processing unit (GPU) resources, making the on-premises deployment of massive LLMs structurally unfeasible.
To resolve this, the project team utilized an open-source small LLM as its foundational model. Through precise domain fine-tuning and rigorous data engineering, the team achieved an impressive risk-rating accuracy rate exceeding 90%. This result not only far surpassed the off-the-shelf, general-purpose LLM's accuracy of over 40%, but also significantly outperformed a larger model with 400% more parameters, which achieved just over 30% accuracy. This breakthrough provides the Client with a powerful, intelligent assistance tool for transaction compliance while validating a new fintech implementation paradigm: replacing massive parameters and high compute costs with compact models and robust fine-tuning.
Data Engineering Breakthroughs: Synthetic Data Augmentation
The original dataset comprised a small set of authentic, anonymized cases — a volume insufficient for effective model fine-tuning. To overcome this, the
- Risk Distribution Matching: Seed data was strictly curated to mirror the exact distribution of the Client's real-world business scenarios—a vast majority of low-risk, a moderate portion of medium-risk, and a small minority of high-risk—ensuring the model's output would not deviate from actual operational realities.
- AI-Powered Data Augmentation: Utilizing generative rewriting strategies, the seed records were significantly expanded into a substantial collection of distinct datasets. This approach preserved core risk patterns while varying the syntactical structure, perfectly balancing data validity with semantic diversity.
To address the class imbalance inherent in datasets dominated by low-risk cases, the team introduced class-weighted loss functions and stratified sampling techniques. These adjustments significantly enhanced the model's sensitivity and precision in identifying high-risk cases.
Overcoming Generative Output Challenges: Task Decoupling and Mandatory Evidence Citation
To mitigate the common generative AI issues of "hallucinations" and context interference when processing complex workflows, the
- Task Decoupling Design: The generation of risk rating, Request for Information (RFI) inquiries, and final reports were decoupled into independent output streams. This architectural shift eliminated the accuracy degradation typically caused by cross-task interference.
- Reasoning Chain Optimization: Initial attempts to inject step-by-step reasoning rationale to improve explainability inadvertently reduced overall accuracy. The team decisively pivoted to a staged decoupling approach with class weighting. Explainability ultimately secured through a comprehensive seven-dimensional scorecard, ensuring transparency without compromising model performance.
Seven-Dimensional Scorecard and Human-in-the-Loop (HITL) Safeguard
The evaluation system also integrated a third-party LLM to conduct blinded A/B testing (LLM-as-a-Judge). By randomizing the output order of fine-tuned versus non-fine-tuned models during the scoring process, the team effectively eliminated positional bias.
For extreme edge cases, the project strictly adhered to HITL mechanism. The Client's business experts retain final decision-making authority, ensuring every review conclusion remains entirely traceable and auditable, fully satisfying the regulatory department's rigorous regulatory requirements for evidence traceability.
Client Feedback and Strategic Outlook
The Client's business units have highly praised the model's efficacy, explainability, and deployment feasibility. By requiring only the deployment of a small LLM under strict resource constraints, the project drastically lowers the barrier to entry for actual production deployment. In a formal letter of commendation, the Client emphasized that since the project's launch this March, the
Throughout the initiative, the
The success of this initiative solidifies
Furthermore,
- Loan Approval Assistance: Automated risk assessment and credit recommendation generation based on the multidimensional data of corporate and individual borrowers.
- Credit Monitoring: Continuous risk monitoring and early-warning detection for existing credit asset portfolios.
- User Behavior Analysis: Identification of anomalous transaction patterns and potential compliance risks.
This scalable methodology will also serve as the foundation for
Mr.
About CLPS Incorporation
CLPS Incorporation (NASDAQ: CLPS), established in 2005 and headquartered in Hong Kong, is at the forefront of driving digital transformation and optimizing operational efficiency across industries through innovations in artificial intelligence, cloud computing, and big data. Our diverse business lines span sectors including fintech, payment and credit services, e-commerce, education and study abroad programs, and global tourism integrated with transportation services. Operating across 10 countries worldwide, with strategic regional hubs in Shanghai (mainland China), Singapore (Southeast Asia), and California (North America), and supported by subsidiaries in Japan and the UAE, we provide a robust global service network that empowers legacy industries to evolve into data-driven, intelligent ecosystems. For further information regarding the Company, please visit: https://ir.clpsglobal.com/, or follow CLPS on Facebook, Instagram, LinkedIn, X, and YouTube.
Forward-Looking Statements
Certain of the statements made in this press release are "forward-looking statements" within the meaning and protections of Section 27A of the Securities Act of 1933, as amended, and Section 21E of the Securities Exchange Act of 1934, as amended. Forward-looking statements include statements with respect to the Company's beliefs, plans, objectives, goals, expectations, anticipations, assumptions, estimates, intentions, and future performance. Known and unknown risks, uncertainties and other factors, which may be beyond the Company's control, may cause the actual results and performance of the Company to be materially different from such forward-looking statements. All such statements attributable to us are expressly qualified in their entirety by this cautionary notice, including, without limitation, those risks and uncertainties related to the Company's expectations of the Company's future growth, deployment in the AI technology sector, performance and results of operations, the Company's ability to capitalize on various commercial, M&A, technology and other related opportunities and initiatives, as well as the risks and uncertainties described in the Company's most recently filed SEC reports and filings. Such reports are available upon request from the Company, or from the Securities and Exchange Commission, including through the SEC's Internet website at http://www.sec.gov. We have no obligation and do not undertake to update, revise or correct any of the forward-looking statements after the date hereof, or after the respective dates on which any such statements otherwise are made.
Contact:
CLPS Incorporation
Rhon Galicha
Investor Relations Office
Phone: +86-182-2192-5378
Email: ir@clpsglobal.com
View original content:SOURCE CLPS