AI PB: Generative Agent on 24×H100 GPUs Delivers Personalized, Compliant Investment Insights

The demand for personalised financial advice is growing, yet delivering trustworthy and compliant insights remains a significant challenge. Daewoo Park, Suho Park, and Inseok Hong, from Shinhan Securities, along with their colleagues, address this need by presenting AI PB, a generative agent designed to proactively offer tailored investment strategies. Unlike conventional chatbots that simply respond to questions, AI PB generates grounded, user-specific recommendations by intelligently combining internal data with external information, all while adhering to strict financial regulations. The team demonstrates that this system, built on a robust infrastructure and incorporating layered safety measures, can deliver reliable and insightful financial guidance in a high-stakes environment, representing a substantial step towards accessible and trustworthy AI-powered investment tools.

Unlike conventional chatbots, AI PB proactively generates grounded and compliant insights, achieved through a component-based orchestration layer that deterministically routes requests between internal and external large language models based on data sensitivity. The system employs a hybrid retrieval pipeline, combining OpenSearch-based sparse retrieval, NMIXX dense retrieval, and domain ontology-based query expansion, to ensure answers are verifiable against enterprise data. Experiments reveal a 30% reduction in hallucination relative to standard prompting methods, achieved by requiring each generated statement to contain at least one reference token.

AI PB continuously generates 22 types of daily insights per user, ranked by a three-layer hybrid recommender system designed to balance relevance, diversity, and novelty. A/B testing demonstrates that this approach increased daily feed engagement by 18% and reduced repetitive content by 23% compared to rule-based baselines. Performance metrics demonstrate an average pre-generation latency of 5. 9 seconds per item and a p95 chat latency under 13. 9 seconds, while maintaining a guard rejection rate below 1.

8%. Human quality assurance reviews, with an inter-rater agreement of κ = 0. 78, recorded average scores of 91. 2% for factuality, 98. 4% for safety, and 85. 7% for alignment. These results confirm the system’s ability to deliver trustworthy AI insights in a high-stakes financial setting.

Trustworthy AI for Financial Insight Generation

AI PB represents a significant advancement in the deployment of generative artificial intelligence within a highly regulated financial environment. Researchers successfully created and deployed a production-scale system capable of proactively generating personalized and compliant investment insights. This achievement stems from a novel combination of deterministic routing of information, a hybrid retrieval system drawing on both broad and finance-specific data, and a multi-stage recommendation mechanism that blends established rules with advanced behavioral modeling. The system demonstrably delivers trustworthy AI insights while maintaining factual reliability, a crucial requirement for high-stakes financial applications.

Importantly, the architecture prioritizes auditability through transparent routing logs, enabling regulators to trace the origin and reasoning behind generated insights. The team acknowledges the need for continuous human oversight to adapt to evolving regulations and refine the system’s performance. Future work may focus on extending this blueprint to other regulated industries, such as banking and healthcare, where similar challenges exist in balancing innovation with compliance and trust.