A multi-agent rag system for auditable token distribution

Har, Sze Hao (2025) A multi-agent rag system for auditable token distribution. Final Year Project, UTAR.

PDF Download (5Mb)

Abstract

This project develops a blockchain-based subsidy program delivered as a web application to enhance the efficiency and transparency of government subsidy distribution. By leveraging blockchain technology, the system ensures secure and auditable allocation of funds, thereby strengthening public trust in subsidy mechanisms. The research methodology involves a review of existing blockchain applications in the public sector, followed by phased development using Visual Studio Code, Supabase, and the Ethereum Sepolia Network. The novelty of this work lies in its integration of multiple innovations. First, an ERC-20 token, MMYRC (Mock Malaysia Ringgit Coin), is created and airdropped to eligible citizens to demonstrate that the entire distribution process can be conducted on-chain. Second, eligibility is determined through a rule-based scoring matrix supported by Retrieval-Augmented Generation (RAG), providing decisions that are both transparent and explainable from government datasets. Third, lightweight AI agents (smolagents) implement a dual-analysis framework that combines the RAG-based approach with a formula-driven burden score, allowing for systematic comparison between interpretability and flexibility in eligibility determination. Finally, zero-knowledge proofs (ZKPs) are explored as a privacy-preserving mechanism, enabling citizens to prove their income bracket (B40, M40, T20) without disclosing precise income data. A mock trusted setup with the Inland Revenue Board of Malaysia is used to verify digitally signed income information, ensuring authenticity while preserving privacy. The expected outcome is a fully functional prototype web application that reduces administrative overhead, improves the auditability of subsidy allocation, and enhances the effectiveness and public legitimacy of government subsidy programs.

Actions (login required)