A MULTI-AGENT-BASED FINANCIAL ASSISTANT APPLICATION

Abstract

Financial literacy in Vietnam has not kept pace with the rapid rise of digital banking and e-wallet usage. Most consumer-facing tools still rely on static content, shallow dashboards, or rule-based chatbots, and many AI assistants remain English-first handling Vietnamese phrasing and local financial terminology poorly. The result is a practical gap: users who want simple, natural-language answers about their own spending trends, budget drift, and savings feasibility often cannot get responses that are both understandable and grounded in their actual transaction history. This thesis presents FinSight, a Vietnamese-focused personal finance and financial education platform designed to close that gap by combining core PFM features (transaction tracking, budgets, goals) with a conversational assistant that explains concepts and analyzes user-specific data. The system uses a service-based architecture with multi-agent orchestration for budgeting, education, reporting, and market-related requests. It grounds responses through a retrieval-augmented pipeline built on curated Vietnamese financial content, and it supports Vietnamese text-to-SQL to translate user questions into constrained, parameterized, read-only database queries with strict user isolation. The work contributes a reference architecture and an implemented prototype, and evaluates key components including intent routing, retrieval quality, and safe query execution, demonstrating that practical, Vietnamese-native, data-grounded financial assistance is feasible with current LLM tooling.