Agentic Data Analyst
PlannedAn AI agent that accepts natural language questions about a dataset and autonomously writes SQL queries, runs pandas operations, generates charts, and explains the results. Built with LangGraph and Streamlit.
Overview
An AI agent that bridges natural language and data — ask a question in plain English, get back SQL results, visualizations, and explanations.
Problem
Querying data requires knowing SQL or pandas. This project removes that barrier by letting an LLM decide how to answer a question, using tools to retrieve and visualize the data.
Approach
LangGraph ReAct agent with three tools: SQL query execution, pandas operations, and chart generation. The agent first fetches the schema, then plans and executes its approach, then explains the result. Streamlit provides the chat interface.
Tech Stack
- LangGraph — agent state machine orchestration
- LangChain — tool definitions and LLM binding
- OpenAI gpt-4o-mini — language model
- SQLite — data store (from Project 1)
- Streamlit — chat UI
Key Decisions
- LangGraph over plain chains because agents that use tools need explicit state and conditional routing — not a linear pipeline
- Tool-level validation restricts SQL to SELECT only — the agent can’t modify data
- Schema fetch as the first tool call ensures the agent always knows the data structure before writing queries
What I’d Do Differently
Add LangGraph’s MemorySaver checkpointer for persistent conversation memory across sessions, so the agent remembers what was discussed earlier in the conversation.