Back to All Projects
Featured Project 0 stars

AI Data Analyst — Multi-Agent Pipeline

Autonomous multi-agent data analysis platform that transforms CSV data into validated insights, visualizations, interactive reports, and report-grounded Q&A.

PythonLangGraphLangChainFastAPIPandasnumpySweetvizPydanticOpenAI-compatible LLMsGroqGeminiLangSmithpytest

Detailed Overview

An autonomous AI Data Analyst that takes a raw CSV file through a complete analysis pipeline. The system profiles the dataset, plans analytical tasks, executes Python analysis in a security-controlled sandbox, validates numerical evidence, generates business insights and recommendations, and produces an executive-ready interactive report. It also provides a report-grounded chatbot for asking questions about the analyzed data. The platform includes both a modern React/TanStack dashboard and a Streamlit interface, with a FastAPI backend and LangGraph-based orchestration. A key focus of the project is reliability: analytical numbers are deterministically validated and verified by recomputing them from the source CSV before they are passed to the LLM. The system also supports provider failover, deterministic pandas fallbacks, execution limits, AST-based security controls, caching, observability, and graceful degradation when external LLM services are unavailable.

Problem Statement

Traditional data analysis requires analysts to manually inspect datasets, write Python code, create visualizations, validate statistics, and prepare reports. LLM-based data-analysis tools can automate much of this work but may generate incorrect calculations, hallucinated insights, or unsafe code execution. This project addresses both the automation and reliability problems by combining multi-agent orchestration with sandboxed execution and deterministic evidence verification.

Technical Solution

A LangGraph-powered multi-agent pipeline divides the work into three stages: Profiler, Analysis, and Insight/Report. The Profiler performs dataset profiling and EDA; the Analysis agent plans and executes analytical tasks using sandboxed Python and a reflection/QA loop; and the Insight agent validates results, verifies numbers by recomputation, generates insights and recommendations, and compiles the final report. A report-grounded chatbot allows users to ask questions using only verified pipeline results.

Architecture & Data Flow

CSV Upload → Profiler Agent → Analysis Planner → Sandboxed Python Executor → Reflector/QA Loop → Deterministic Validation → Evidence Extraction & Verify-by-Recompute → LLM Insights & Recommendations → Jinja2 Report Generator → Sweetviz Interactive Report + Summary Report → Report-Grounded Chatbot Shared AgentState is used across the pipeline to coordinate profiling results, analysis plans, execution results, validation reports, insights, recommendations, reports, and logs. The backend exposes FastAPI endpoints while the React/TanStack and Streamlit applications provide user interfaces.