Data Scientist Agent

A LangGraph-powered, multi-agent data science assistant with a Flask API and lightweight front-end. The app plans, executes, and summarizes an end-to-end workflow (loading, cleaning, EDA, statistical tests, wrangling, prep, viz, modeling, insight summary) and generates a downloadable Jupyter notebook.

Features

• Planning agent defines the workflow steps.
• Specialized sub-agents for loading, cleaning, EDA, statistical analysis, data wrangling, preparation, visualization, predictive modeling, and summary.
• Flask API to trigger runs, fetch health, and download notebooks.
• Front-end to upload CSVs, view the plan, inspect outputs, and grab the generated notebook.
• Notebook generation with markdown/code cells per step for reproducibility.

Quickstart

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run the app

flask --app app run --host 0.0.0.0 --port 5000

Open http://localhost:5000 to use the UI.

API

• GET /api/health – liveness probe.
• POST /api/run – run the full flow. Form-data fields: instructions (text), optional file (CSV upload).
• GET /api/notebook/<filename> – download the generated notebook.

See API_DOC.md for full details and payload examples.

How it works

The orchestrator builds a LangGraph with nodes representing the planning agent plus sub-agents for each data science task. Each node updates shared state, logs outputs, and appends notebook cells. After execution, the service returns:

• Plan – ordered steps.
• Outputs – structured text per step.
• Summary – concise highlights.
• Notebook – downloadable artifact covering the full run.

If no dataset is uploaded, the system uses a small synthetic dataset to keep the demo runnable.

Development notes

• Outputs are persisted under uploads/, notebooks/, and static/ (for generated plots).
• Visualization uses a headless Matplotlib backend.
• Predictive modeling defaults to LogisticRegression; if the dataset lacks a target column, a synthetic one is derived from numeric features to keep training functional. Target inference respects common names (target, label, outcome, y) and low-cardinality columns before synthesizing.
• Visualization gracefully skips plots when no numeric columns are present and can render histograms, boxplots, scatter plots, correlation heatmaps, scatter matrices, and line plots based on the requested instructions.