Build a “Chat With Your Data” AI Agent in Python + Claude

Build a “Chat With Your Data” AI agent using Python, Claude, DuckDB, and SQL — then see why the simplest approach can give you confidently wrong answers. In this video, we build a working chat-with-your-data system from scratch and test three approaches: sending the entire dataset to the LLM, sending only the database schema, and adding sample rows to give the model better context. We’ll use real NYC energy audit data with roughly 10,000 rows and 2,000 columns. You’ll see how: • Sending all your data quickly exceeds LLM context limits • Claude can generate SQL directly from a database schema • Misleading column names can cause plausible but incorrect answers • Adding a small amount of context can dramatically improve the result This is the first video in a series on building a more reliable AI agent for data analysis. We’ll progressively add data dictionaries, agentic loops, RAG, skills, guardrails, and eventually a complete chat UI. The goal isn’t just to build something that can “chat with your data.” It’s to build an AI data analyst you can actually trust. Anthropic article: https://claude.com/blog/how-anthropic... NYC Energy Audit Data: https://data.cityofnewyork.us/Environ... Want to work together? matt@gosolucia.com https://github.com/MattBrown88/ai-dat... MatthewLBrown.com #AI #Python #Claude #DataAnalytics #LLM