Financial Analyst AI Agent System: An Intelligent Stock Analysis Platform with Multi-Expert Perspectives

The Financial Analyst AI Agent System introduced in this article is a multi-agent AI system built on Google ADK. It integrates the analytical frameworks of three investment masters—Warren Buffett, Cathie Wood, and Greg Abel—provides real-time stock analysis services via FastAPI, supports multiple investment style perspectives, and offers users multi-dimensional professional-level stock analysis.

In the fintech field, artificial intelligence is reshaping traditional investment analysis processes. Traditional analysis relies on analysts' personal experience and single perspectives, while modern investors need multi-dimensional, real-time, and customizable services. This open-source project emerged as a response. Its core innovation is transforming "celebrity investment perspectives" into an executable AI agent system. Users can examine stocks from different perspectives such as Warren Buffett (value investment), Cathie Wood (disruptive innovation), and Greg Abel (operational excellence), making investment concepts actionable.

The system is built on Google ADK and adopts a layered agent architecture:

1. Planning Agent Layer: Acts as an entry coordinator, understanding user questions, identifying stock codes, and routing tasks to appropriate expert agents;
2. Expert Agent Layer: Includes three expert models—Warren Buffett Agent (value investment, focusing on moats, long-term advantages, etc.), Cathie Wood Agent (disruptive innovation, focusing on technological changes and growth prospects), Greg Abel Agent (operational excellence, focusing on efficiency and capital allocation);
3. Tool and Data Layer: Integrates the Finnhub API to provide real-time financial data, including stock tools (quotes, K-lines), news tools (company and market news), and fundamental tools (overviews, financial data).

The core implementations of the system include:

1. Celebrity Skill Embedding: Encoding investment masters' analytical frameworks into structured skill modules (including core principles, question lists, key indicators, decision weights), which are closely integrated with tool calls and data acquisition;
2. LiteLLM and Multi-Model Support: Communicating with the MiniMax M2.7-highspeed model via the LiteLLM agent layer to achieve model agnosticism, facilitating switching between LLM providers;
3. FastAPI REST Interfaces: Providing endpoints such as /analyze (core analysis), /search (data search), /agents (expert list), and /clear-session (session management), supporting API key authentication and health checks.

Typical user workflow:

1. The user asks a question in natural language (e.g., "What does Tesla look like from Warren Buffett's perspective?");
2. The planning agent parses the question, identifies the stock code, and determines the analysis style;
3. Activates the corresponding expert agent(s) (single or multiple);
4. The expert agent calls Finnhub tools to obtain real-time data;
5. The LLM generates analysis based on the data and skills;
6. Returns a structured JSON response containing conclusions from each perspective. Supports single or multi-perspective comparative analysis (style: "all" mode).

The project's highlights include:

1. Actionable Investment Concepts: Transforming abstract investment concepts (e.g., Warren Buffett's moat theory) into AI agents' behavioral patterns, distinguishing it from traditional quantitative models;
2. Multi-Agent Collaboration Mode: Parallel expert mode where each agent independently analyzes the same target, reflecting differences in various investment philosophies;
3. Integration of Real-Time Data and LLM: Dynamically obtaining the latest data via tool calls to address the limitations of LLM's static knowledge and ensure the timeliness of analysis.

The project has the following limitations and improvement directions:

1. Data Dependence: Relies on the Finnhub API, and some market or non-US stock data may be incomplete;
2. Skill Depth: Current celebrity skills are relatively simplified; need to improve the encoding of masters' intuition and experience;
3. Backtesting Validation: Lacks historical backtesting functionality; needs to be supplemented to evaluate strategy effectiveness;
4. Risk Disclosure: Need to clearly state that AI analysis is not investment advice and add a risk disclosure mechanism.

Summary: This project has solid technology and innovative design. It successfully applies the multi-agent architecture to investment analysis and provides a unique celebrity perspective analysis experience. Although it requires more validation and risk control, its AI engineering exploration has reference value. Insights: Lessons for developers include agent specialization, structured knowledge embedding, real-time data source integration, and user-controlled analysis perspectives. Such systems are expected to play a greater role in the financial analysis field.