Scholarlink

Risk assessment in construction supply chain finance is hampered by intricate shareholder relationships, multilingual and semi-structured documentation, and rapidly changing market conditions. We present ARO (Agentic Risk Orchestrator), a novel multi-agent system powered by large language models (DeepSeek R1 [7], LLaMA 3 [8] and Fin-LLaMA [12]) that orchestrates specialized agents through a central planner to perform document understanding, financial analysis, market intelligence gathering, and report synthesis. ARO introduces a dynamic, stakeholder-aware risk tree to fuse internal financial signals with external market and regulatory signals, and computes an aggregate risk score spanning financial, market, operational, and sentiment components. In a real-world Hong Kong SME contractor case study, ARO processed multi-year audits, bank statements, and public news to produce an explainable risk report in minutes, compared to hours for manual workflows. We contribute: 1) a domain-tailored multi-agent architecture for construction finance; 2) a formal coordination framework and weight-calibrated risk model with auditability and confidence scoring; 3) integration of real-time market intelligence into a dynamic risk tree; and 4) a complete case study illustrating decision support for underwriting. We position ARO relative to recent multi-agent credit systems such as MASCA [9], highlighting domain specialization, multilingual OCR, and stakeholder-graph analytics as key differentiators. We release formal specifications to support reproducibility and future benchmarking.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ARO: Multi-agent LLM System for Construction Supply Chain Finance Risk Assessment

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pritom Rajkhowa,
  • Eddie Lin,
  • Ryan Wong,
  • Jestyn Khoo,
  • Abhijit Baishya,
  • Mingzhen Jiang

摘要

Risk assessment in construction supply chain finance is hampered by intricate shareholder relationships, multilingual and semi-structured documentation, and rapidly changing market conditions. We present ARO (Agentic Risk Orchestrator), a novel multi-agent system powered by large language models (DeepSeek R1 [7], LLaMA 3 [8] and Fin-LLaMA [12]) that orchestrates specialized agents through a central planner to perform document understanding, financial analysis, market intelligence gathering, and report synthesis. ARO introduces a dynamic, stakeholder-aware risk tree to fuse internal financial signals with external market and regulatory signals, and computes an aggregate risk score spanning financial, market, operational, and sentiment components. In a real-world Hong Kong SME contractor case study, ARO processed multi-year audits, bank statements, and public news to produce an explainable risk report in minutes, compared to hours for manual workflows. We contribute: 1) a domain-tailored multi-agent architecture for construction finance; 2) a formal coordination framework and weight-calibrated risk model with auditability and confidence scoring; 3) integration of real-time market intelligence into a dynamic risk tree; and 4) a complete case study illustrating decision support for underwriting. We position ARO relative to recent multi-agent credit systems such as MASCA [9], highlighting domain specialization, multilingual OCR, and stakeholder-graph analytics as key differentiators. We release formal specifications to support reproducibility and future benchmarking.