TLDR: AI PB is a production-scale generative AI agent deployed by Shinhan Securities for retail finance. It proactively generates personalized, grounded, and compliant investment insights using a component-based orchestration layer for deterministic routing, a hybrid retrieval pipeline for factual evidence, and a multi-stage recommendation mechanism for personalization. Operating fully on-premises under strict Korean financial regulations, the system leverages Docker Swarm and vLLM on NVIDIA H100 GPUs. It features a dual-interface (Today Feed and Dialogue View) and incorporates Shinhan-Guard for safety and PII protection. Evaluations show high factuality and compliance, demonstrating a blueprint for trustworthy AI in high-stakes finance.
Shinhan Securities has introduced AI PB, an innovative generative AI agent designed to provide personalized investment insights to retail finance users. This system moves beyond traditional chatbots by proactively generating grounded, compliant, and user-specific financial advice, addressing critical challenges like hallucination and regulatory risks in the financial sector.
AI PB operates with a core philosophy of “ground-first generation,” meaning all its textual outputs are built upon verified enterprise data. This ensures the reliability and trustworthiness of the information provided. A key feature is its deterministic routing system, which intelligently directs queries to either internal or external large language models (LLMs) based on data sensitivity, guaranteeing that personal identifiable information (PII) never leaves the institution.
The system offers a dual-interface experience to users: the “Today Feed” and the “Dialogue View.” The Today Feed is a dynamic dashboard that delivers pre-generated daily briefings, including market updates, sector narratives, and personalized portfolio summaries. The Dialogue View, on the other hand, allows for interactive, multi-turn conversations, providing contextual grounding and visual elements for a comprehensive advisory experience. This design enables a seamless transition between passively consuming information and actively exploring financial data.
Underpinning AI PB’s functionality is a robust architecture comprising a Single-Agent System and a Recommendation System. The Single-Agent System integrates 20 components, each representing a high-level workflow like “analyze my portfolio” or “summarize recent disclosures.” These components are built from 48 modules, which are executable tools that interface with various internal data systems such as AltiBaseDB, OracleDB, PostgreSQL, OpenSearch, and legacy TR systems. This modular design enhances maintainability and traceability.
Safety and compliance are paramount for AI PB. All inputs and outputs are rigorously filtered through Shinhan-Guard, a specialized derivative of Llama Guard 3, fine-tuned for the Korean financial context. This guard mechanism ensures that no policy-violating or harmful text reaches end-users, with rejections triggering fallback to safe templates. The system’s routing policy is designed to prevent PII egress, ensuring that sensitive data is always handled by internal models.
To combat hallucination, AI PB employs a hybrid retrieval pipeline. This pipeline combines OpenSearch for keyword matching, the finance-domain specific NMIXX embedding model for semantic relevance, and query expansion using domain ontologies. Retrieved passages are then structured into evidence templates, providing verifiable context for the generative models. A post-generation validator further ensures that every statement is linked to a reference, significantly reducing hallucination.
Personalization is achieved through a multi-stage recommendation mechanism. AI PB proactively generates 22 types of daily insights per user, which are then ranked by a three-layer hybrid recommender. This recommender balances relevance, diversity, and novelty using rule-based prioritization, sequential behavioral modeling based on user interaction patterns, and a contextual bandit layer that adjusts rankings in real-time based on user feedback. This approach has shown to increase daily feed engagement and reduce repetitive content.
The entire system operates on-premises within a Docker Swarm cluster, utilizing vLLM for efficient inference across 24 NVIDIA H100 GPUs. This on-premise deployment ensures full data sovereignty, adhering to strict Korean financial regulations. The generative workloads are handled by the Qwen3-32B ORPO-aligned model, with LoRA+ORPO tuning applied to maintain stylistic and factual consistency. The NMIXX embedding model, specialized for financial semantics, further enhances retrieval accuracy.
Evaluations have demonstrated AI PB’s effectiveness, with human QA showing over 90% factuality and near-perfect compliance. The system maintains high recall on harmful or privacy-sensitive content while preserving fluent responses. The proactive pre-generation of insights helps amortize the latency cost associated with extensive retrieval, ensuring responsiveness. The component-level routing also creates transparent audit trails, crucial for regulatory compliance.
Also Read:
- HCLA: A Human-Centered Approach to Detecting Anomalous Digital Asset Transactions
- Fetch.ai Unveils a Unified Architecture for Intelligent Multi-Agent Systems
AI PB represents a significant step forward in deploying generative AI in regulated industries, offering a blueprint for creating trustworthy and personalized AI insights at an enterprise scale. For more details, you can refer to the research paper.
