Projects
Virtual Buffett: AI Investment Advisor
This project introduces a custom-built AI agent inspired by Warren Buffett’s investing principles. Named “Virtual Buffett,” the assistant integrates LangChain, LangGraph, Milvus vector search, and OpenAI’s GPT-4o-mini. Developed for the MSDS 442 course at Northwestern University, the agent retrieves information from Buffett’s shareholder letters and offers context-aware investment insights using a Buffett-style persona.
The assistant blends long-term memory (Milvus), real-time tools (like FRED and Tavily), and a custom persona to emulate Buffett’s plainspoken wisdom and investment philosophy, making it suitable for financial education, decision support, and investor engagement use cases.
⸻
Features
- Buffett Persona: Customized prompt engineering using 15 distilled Buffett principles.
- Vector Search with Milvus: Embeds and indexes all shareholder letters from 1977–2023.
- Tool Use: Includes FRED API for economic data, Tavily for web results, and FMCSA lookup for fundamentals.
- Long-Term Memory: Stores structured conversation history using SQLite + vector metadata.
- Modular Architecture: Built using LangGraph to control reasoning over tools, memory, and output.
- Progress Visualization: Includes dynamic progress bars and conversational threading with memory awareness.
⸻
Key Insight
Virtual Buffett demonstrates how an LLM-based agent can emulate human investment reasoning by retrieving real-world evidence, applying structured principles, and offering interpretable, long-term-focused responses. The project shows that AI personas built on domain-specific corpora can deliver personalized, expert-level guidance, without hallucination.