LLM-Assisted EDA
Upload any CSV and instantly receive Gemini-generated insights, visualizations, and reports — with automated profiling for missing values, duplicates, and IQR-based outliers.
Overview
An automated exploratory data analysis tool that uses Google's Gemini AI to transform CSV datasets into insights, visualizations, and reports through a single command. Features smart data profiling that detects missing values, duplicates, and outliers via IQR methodology, AI-driven insights with evidence-based observations citing specific numerical values, and context-aware plot selection (boxplots for outliers, bar charts for categorical data). The modular pipeline handles analysis statistics, AI integration, plot generation, and report orchestration.
Problem Statement & Approach
Exploratory data analysis is repetitive and manual — profiling a new dataset, choosing the right plots, and writing up findings all take time before any real insight emerges.
Approach: Automate the first pass: point the tool at a CSV and let it profile the data, pick context-appropriate charts, and use Gemini to generate evidence-based, numerically grounded observations and a report in a single command.
System Architecture
A modular pipeline with four stages — analysis statistics (profiling, missing values, duplicates, IQR outliers), AI integration (Gemini insight generation), plot generation (context-aware chart selection), and report orchestration that assembles the final output.
Key Features
- Transforms any CSV into insights, visualizations, and reports in one command
- Smart profiling for missing values, duplicates, and IQR-based outliers
- Evidence-based AI insights citing specific numerical values
- Context-aware plot selection (boxplots for outliers, bar charts for categoricals)
- Modular pipeline: statistics, AI integration, plotting, and report orchestration
Technical Stack
Language
AI
Data
Visualization
Deployment
Runs as a command-line tool — a single command turns a CSV into a full insights report.
Challenges & Solutions
Challenge: LLM-generated insights can drift into vague or unsupported claims.
Solution: Evidence-based prompting requires each observation to cite specific numerical values from the profiled data.
Challenge: The right visualization depends on each column's type and distribution.
Solution: Context-aware plot selection chooses boxplots for outliers, bar charts for categoricals, and so on automatically.
Improvements
- Support for databases and Parquet beyond CSV
- Interactive HTML reports
- A model-agnostic backend to swap in other LLMs
