FinCom (Financial Committee) is a supervisor-based multi-agent framework designed to support structured, committee-style financial decision-making. Built on LangGraph, the system orchestrates three domain-specialized agents—Research, Quantitative, and Risk Management—each powered by large language models with ReAct-based reasoning and constrained tool access. These agents collaborate under a central Supervisor to analyze financial queries from multiple perspectives and synthesize unified, actionable recommendations. To address limitations of traditional single-agent systems, FinCom introduces a flexible coordination framework with multiple architecture modes, including centralized supervision, decentralized agent handoff, and a structured committee workflow. In its default configuration, the system employs a two-stage process: parallel independent analysis followed by sequential deliberation using a Disagree-or-Commit (DoC) protocol, encouraging explicit critique and reducing conformity bias in multi-agent reasoning. The system is implemented as an end-to-end application with a React-based frontend and a LangGraph backend, supporting real-time, multi-agent interaction through a chat interface. Evaluation is conducted using LangSmith’s LLM-as-a-Judge pipeline, combined with both benchmark-based and workflow-oriented tasks to assess reasoning quality, coherence, and robustness. FinCom demonstrates how multi-agent coordination, tool-augmented reasoning, and structured deliberation can be integrated into a practical financial analysis system. The project highlights key challenges in agent orchestration, evaluation design, and user-facing transparency, providing insights into the development of scalable, explainable AI systems for complex decision-making domains.
Mentor: David Ye
Project poster (PDF)
