Financial AI Agent Quickstart

Table of Contents

• Financial AI Agent Quickstart
  • Table of Contents
  • About The Project
  • Prerequisites
  • Environment Setup
  • Installation
  • Directory Structure
  • Setup
  • Input Data
  • Usage Examples
    • Reflexive Metacognitive Financial Advisor with LangGraph
    • SEC 10-K & 10-Q Filings Analysis with LlamaIndex
    • Stock Data and Sentiment Analysis with Agno
    • FinSight: Metacognitive Multi-Agent Earnings Call Analysis
  • Roadmap
    • ✅ Completed
    • 🚧 In Development
    • 🎯 Future Enhancements
  • Contributing
  • License
  • Acknowledgments

About The Project

This project provides a template and best practice for developing multi-agent AI workflows, leveraging various agentic frameworks, modern LLMs and AI evaluation strategies. It demonstrates:

• Multi-agent orchestration strategies across different frameworks (LangGraph, LlamaIndex, Agno)
• Metacognitive AI agents with self-awareness and responsible decision-making capabilities
• AI Safety & Compliance in financial domains with built-in ethical boundaries and professional referral systems
• Integration with local and cloud LLMs, including SOTA (OpenAI / Google Gemini) and Open Weight models (Llama, Qwen, DeepSeek)
• RAG pipeline for document-based retrieval and analysis
• Evaluation metrics for agent orchestration, tool execution and RAG quality

Prerequisites

Before starting, ensure you have:

• Python 3.11.9
• uv: High-performance Python package manager (uv installation)
• Google Gemini API Key: Provided by the FINOS Team and set as GOOGLE_API_KEY in your environment
• Tavily Search API Key: Get from Tavily and set as TAVILY_SEARCH_API_KEY
• Scaleway GenAI API Key: Provided by the FINOS team and set as SCW_SECRET_KEY
• (Optional) LM Studio for local LLM serving (Download LM Studio)

Environment Setup

Rename the .env.new file in your project root to .env and update with the API Keys provided by the FINOS team.

Installation

Clone the repo and install dependencies:

git clone https://github.com/your_username/agent-quickstart.git
cd agent-quickstart
uv sync

Directory Structure

Agent-QuickStart/
├── .env
├── langgraph_agent_example.ipynb      # Metacognitive financial advisor with LangGraph
├── llama_index_agent_example.ipynb    # SEC filings analysis with LlamaIndex
├── agno_agent_example.ipynb           # Stock data and sentiment analysis with Agno
├── finsight_agent.ipynb               # Multi-agent earnings call analysis with LangGraph
├── main.py
├── data/
│   ├── input/          # Place PDF files here
│   └── output/         # Converted markdown files will be stored here
└── database/
    └── vector_store/   # ChromaDB database will be created here

Setup

• Local LLM (Optional):
  • Install LM Studio
  • Download a compatible model (e.g., qwen3-8b)
  • Start LM Studio server on http://127.0.0.1:1234/v1

Input Data

• Place PDF documents (10K or 10Q reports) to be processed in data/input/
• Converted markdown files will be saved in data/output/

Usage Examples

We have four agent examples, each demonstrating different approaches to multi-agent AI workflows and financial analysis:

Reflexive Metacognitive Financial Advisor with LangGraph

The langgraph_agent_example.ipynb notebook demonstrates an advanced Reflexive Metacognitive Financial Agent that maintains explicit awareness of its financial knowledge boundaries, confidence levels, and ethical responsibilities. This agent showcases:

• Self-Aware Decision Making: The agent analyzes each query against its capabilities before responding
• Multi-Strategy Routing: Chooses between direct reasoning, specialized tool usage, or escalation to professionals
• Built-in AI Safety: Explicit ethical boundaries and regulatory compliance in financial advice
• Professional Referral System: Rather than refusing help, provides constructive guidance on finding qualified professionals
• Scaleway AI Integration: Demonstrates integration with Scaleway's GenAI API infrastructure

The agent asks critical questions like:

• "Do I have sufficient expertise to provide guidance on this financial topic?"
• "Is this query within my designated area of financial knowledge?"
• "Is this a high-stakes financial decision where incorrect advice could cause significant harm?"

This example is particularly valuable for understanding AI safety and responsibility in financial domains, where poor advice can have serious real-world consequences.

SEC 10-K & 10-Q Filings Analysis with LlamaIndex

The llama_index_agent_example.ipynb notebook demonstrates how to:

• Load environment variables
• Convert PDFs to markdown
• Generate vector database
• Query and evaluate RAG pipeline
• Define and run multi-agent workflow
• Leverage DeepEval across the flow execution, with a LLM-as-Judge approach

Stock Data and Sentiment Analysis with Agno

The agno_agent_example.ipynb notebook demonstrates how to:

• Leverage Google Search for web search on recent information to produce sentiment analysis
• Leverage Yahoo Finance for financial data retrieval including historical stock prices and financial ratios
• Define and run multi-agent workflow to produce an analyst report

FinSight: Metacognitive Multi-Agent Earnings Call Analysis

The finsight_agent.ipynb notebook demonstrates an advanced FinSight Agent - a metacognitive multi-agent system designed for comprehensive financial analysis of earnings call transcripts. This system showcases:

• Metacognitive Coordinator: Master agent with self-awareness that interprets queries, plans analysis workflows, and synthesizes findings from specialized expert agents
• Specialized Expert Agents:
  • Sentiment Analysis Agent: Correlates earnings call sentiment with market news and broader sentiment signals
  • Significant Event Detection Agent: Identifies and validates corporate events by cross-referencing official sources (SEC filings, press releases)
  • Volatility Prediction Agent: Predicts stock volatility based on transcript analysis and multimodal features
• Configurable Analysis Framework: Flexible question-based system for extracting structured insights from earnings calls
• Guardrail Tracking: Transparent reporting of safety measures, confidence thresholds, and operating boundaries
• Multi-Modal Integration: Combines transcript analysis, financial news, market data (Yahoo Finance), and SEC filings
• Structured Reporting: Comprehensive reports with confidence scores, source verification, and disclaimers

The FinSight agent demonstrates best practices for building production-ready financial AI systems with:

• Explicit confidence thresholds for each agent
• Operating boundaries and ethical guardrails
• Source verification requirements
• Investment advice prohibition safeguards
• Transparent limitation disclosure

This example is particularly valuable for understanding how to orchestrate multiple specialized agents in a financial domain while maintaining safety, transparency, and accountability.

Roadmap

✅ Completed

• Multi-agent financial analysis workflow (Research → Write → Review)
• Reflexive Metacognitive Financial Agent with LangGraph
• FinSight Multi-Agent Earnings Call Analysis System
• AI Safety & Compliance framework for financial domains
• Multi-strategy routing (Direct reasoning, Tool usage, Professional escalation)
• Self-aware decision making with confidence assessment
• RAG pipeline with financial document processing
• Comprehensive AI evaluation metrics (DeepEval integration)
• Real-time quality assessment during agent execution
• Professional referral and escalation system
• Specialized agents for sentiment analysis, event detection, and volatility prediction

🚧 In Development

• Additional Financial Use Cases
  • ESG (Environmental, Social, Governance) analysis agents
  • Regulatory compliance assessment workflows
  • Financial risk analysis and stress testing
  • Market sentiment analysis from earnings calls

🎯 Future Enhancements

• Advanced Evaluation Strategies

  • Custom financial domain evaluation metrics
  • Bias detection in financial analysis
  • Hallucination detection for financial facts
  • Regulatory compliance validation

• Multi-Modal Analysis

  • Financial chart and graph analysis
  • Integration with financial data APIs (Bloomberg, Reuters)
  • Audio analysis of earnings calls

• Production Features

  • Model performance monitoring and alerts
  • A/B testing framework for agent improvements
  • Audit trail and explainability features
  • Integration with financial data platforms

Contributing

All commits must be signed with a DCO signature. See CONTRIBUTING.md for details.

Signed-off-by: Your Name <your.email@example.com>

Add the -s flag to your git commit to sign automatically.

License

Distributed under the Apache License, Version 2.0. SPDX-License-Identifier: Apache-2.0

Acknowledgments

• LangGraph - State-based agent framework for building metacognitive AI systems
• LlamaIndex - Multi-agent workflow framework and RAG implementation
• DeepEval - AI evaluation metrics and testing framework
• LM Studio - Local LLM serving platform
• Tavily - Web search API for real-time information retrieval
• ChromaDB - Vector database for document embedding and retrieval
• Scaleway - AI infrastructure and GenAI API services