Cathay Financial Holdings is advancing generative AI through its GAIA framework and an AI-as-a-Service strategy, emphasizing open-source small language models for customer intent classification.

A study of models from Meta, TAIDE, TAME, NVIDIA, and OpenAI found that fine-tuned SLMs can reduce reliance on complex prompt engineering and vector retrieval, potentially simplifying architecture and lowering maintenance.

Fine-tuning with finance-domain data and synthetic data improved stability, inference efficiency, and deployment controllability, achieving performance in customer intent classification close to mainstream closed-source LLMs.

Future applications include mortgage balance inquiries, credit card payment assistance, and branch service navigation, enabling intelligent search, service routing, and enhanced customer engagement.

These applications emphasize intelligent search, service routing, and elevated customer engagement within branch and digital channels.

A fully synthetic data approach was used to protect customer privacy, employing techniques such as service-function clustering, single- and multi-intent dataset design, Taiwan-context localization, and keyword expansion to improve local context understanding.

This synthetic data strategy addresses data governance and privacy by avoiding real customer data in training.

Techniques include service-function clustering, single- and multi-intent dataset design, Taiwan-context localization, and keyword expansion to enhance understanding of local finance contexts.

Looking ahead, the company plans to pursue long-context classification, advanced financial document understanding, and cross-scenario AI applications to accelerate innovation and improve customer experience in financial services.

The roadmap also envisions cross-scenario AI applications tailored to finance to push innovation and boost customer-centric service.

Technical architecture integrates NVIDIA tools (NeMo Customizer, NeMo Curator, TensorRT-LLM) with NVIDIA Hopper hardware to support data generation, fine-tuning, inference optimization, and evaluation within the NVIDIA AI ecosystem.

The setup underscores NVIDIA's ecosystem for financial-domain model development and governance, combining specialized software and hardware for end-to-end AI workloads.