Personal Financial AI Agent Assistant

🚧 Proof of Concept - A stateful multi-agent AI system for personalized financial advice using LangGraph, LangChain, and MCP tools.

Overview

A multi-agent AI system that combines specialized financial agents to deliver personalized financial advice. Each agent focuses on a specific domain and collaborates through a cyclical LangGraph orchestration layer, allowing for iterative research and quality review.

🤖 AI Agents

Agent	Role	Status	Capabilities
💼 Investment Agent	Specialist	✅ Active	Portfolio recommendations, market analysis
🔍 Review Agent	Quality Control	✅ Active	Validates outputs, provides feedback, ensures consistency
🎯 Supervisor Agent	Orchestrator	🔄 In Development	Routes queries, manages workflow, coordinates agents
🤖 Generalist Agent	General Assistant	📋 Planned	Handles general queries unrelated to finance, conversational support
📊 Tax Specialist Agent	Specialist	🔄 In Development	Tax optimization, deduction strategies
⚠️ Risk Assessment Agent	Specialist	📋 Planned	Risk evaluation, mitigation strategies
💰 Budget Planner Agent	Specialist	📋 Planned	Expense tracking, budget optimization

Technology: LangGraph • LangChain • Model Context Protocol (MCP) • Python 3.11+

Agent Flow (Powered by LangGraph StateGraph):

1. 🎯 User Query is passed into the initial state.
2. 💼 Investment Agent (Researcher) performs analysis using Search tools and generates a report.
3. 🔍 Review Agent validates the output.
4. Conditional Router:
   • If Reviewer replies with PASS ➡️ Return final response to User.
   • If Reviewer replies with FAIL ➡️ Loop back to the Researcher with feedback for refinement (up to 3 iterations).

🔌 MCP Servers Integration

The system will leverage multiple MCP servers to extend functionality:

• Web Search Server

  • Real-time market data and financial news retrieval
  • General web search for non-financial queries (Generalist Agent)
  • Current: Using LangChain Google Search Wrapper
  • Planned: Migration to dedicated MCP Search server

• File Data Server (planned)

  • Excel/CSV file processing for user-provided portfolio data
  • Historical transaction analysis

• Database Server (planned)

  • Persistent storage for user preferences and transaction history
  • Portfolio tracking across sessions

💻 User Interface

• Current: Streamlit-based chat UI with real-time agent interaction ✅
• Next Phase: Migration to React/Next.js for production-ready web application (planned)

🏗️ Architecture

Current (v1.5 - LangGraph State Orchestration)

The system currently uses a cyclic graph architecture to ensure high-quality outputs.

Flow:

1. User submits financial query via Streamlit chat interface.
2. The StateGraph initializes the state dictionary.
3. Research Node fetches real-time data and drafts a plan.
4. Review Node audits the plan.
5. If the plan fails the audit, it is routed back to the Research Node with constructive feedback.
6. Final verified response returned to user in chat UI.

Planned (v2.0 - Supervisor Routing + MCP)

Enhanced Multi-Agent System:

• 🎯 Supervisor/Orchestrator - Routes queries, manages workflow
• 🤖 Generalist Agent - General conversational queries
• 💼 Specialist Agents - Investment, Tax, Risk, Budget
• 🔍 Review Agent - Quality control & feedback

🚀 Quick Start

Prerequisites

• uv - Python package and project manager
• Python 3.11+
• .env file containing your GEMINI_API_KEY, GOOGLE_API_KEY, and GOOGLE_CSE_ID

Installation

# Install uv
curl -LsSf [https://astral.sh/uv/install.sh](https://astral.sh/uv/install.sh) | sh

# Install dependencies
uv sync

# Run the Streamlit app
uv run streamlit run streamlit.py