TLDR: Shinhan Securities has introduced ‘AI PB,’ a groundbreaking generative AI agent operating on 24 NVIDIA H100 GPUs. This system proactively delivers personalized and compliant investment insights to retail finance customers, addressing the critical challenges of trustworthiness, data privacy, and regulatory adherence in AI-driven financial advice.
In a significant leap for financial technology, Shinhan Securities, in collaboration with researchers including Daewoo Park, Suho Park, and Inseok Hong, has unveiled ‘AI PB’ (AI Private Banker). This production-scale generative agent is designed to provide proactive, personalized, and fully compliant investment insights to retail finance customers, marking a new era in AI-driven financial advisory services.
The development comes amidst a surging demand for tailored financial advice, which has historically been difficult to scale while maintaining trustworthiness and strict regulatory compliance. Traditional large language model (LLM) deployments often grapple with issues like hallucination, unverifiable reasoning, and the inherent risks of handling sensitive financial data, especially under stringent privacy regimes where personal information cannot leave institutional boundaries.
AI PB directly confronts these challenges through a meticulously engineered architecture. Unlike reactive chatbots, this agent proactively generates insights, integrating a sophisticated component-based orchestration layer that intelligently routes queries. This routing mechanism deterministically selects between internal and external LLMs based on data sensitivity, ensuring zero Personally Identifiable Information (PII) egress. All textual outputs are constructed from verified enterprise data, adhering to a ‘ground-first generation’ principle to guarantee factual reliability.
The system operates entirely on-premises within Shinhan Securities’ data center, a crucial design choice driven by strict Korean financial regulations concerning data residency and algorithm transparency. The robust infrastructure leverages Docker Swarm and vLLM for efficient inference across 24 NVIDIA H100 GPUs. Specifically, eight GPUs are dedicated to guard, embedding, and reranking tasks, while the remaining sixteen handle generative workloads, primarily based on the Qwen3-32B ORPO-aligned model. Model alignment is further enhanced through LoRA+ORPO tuning on 10,000 synthetic, expert-reviewed outputs to ensure stylistic and factual consistency, as well as regulatory compliance.
AI PB offers a dual-interface experience: a ‘Today Feed’ for pre-generated daily briefings and a ‘Dialogue View’ for interactive, multi-turn conversations. Its architecture comprises a Single-Agent System, integrating 20 components (18 informational, 2 personal-asset analytic) built on 48 data modules that interface with internal systems like AltiBaseDB, OracleDB, and OpenSearch. A Recommendation System then ranks and surfaces these pre-generated insights for the ‘Today Feed.’
Also Read:
- ComplyAdvantage Unveils Mesh: An AI-Powered Platform to Combat Surging Financial Crime
- FINRA Expands AI Integration for Enhanced Market Surveillance and Risk Oversight
This innovative approach, which combines deterministic routing, a hybrid retrieval pipeline utilizing OpenSearch and a finance-domain embedding model, and a multi-stage recommendation mechanism, demonstrates that grounded, compliant, and personalized generative agents can be successfully deployed at an enterprise scale in highly regulated industries. The creators believe AI PB offers a blueprint for similar applications in other sensitive sectors such as banking and healthcare, setting a new standard for trustworthy AI in high-stakes financial environments.
