Autonomous ML Agent 🤖

An end-to-end autonomous machine learning agent that ingests datasets, evaluates multiple models, and recommends the best predictor based on evaluation metrics.
This tool is designed for developers who want a quick, automated way to test and compare ML models without writing boilerplate code.

🎥 Demo Video

---

🚀 Features

• Automated data preprocessing (handling missing values, scaling, encoding).
• Supports multiple machine learning models out of the box.
• Evaluation and ranking of models using standard metrics (accuracy, F1-score, RMSE, etc.).
• Visualizations for performance comparison.
• Modular design → easily extendable with new models or metrics.

---

🛠️ Tech Stack

• Python 3.10+
• scikit-learn – training & evaluation
• pandas / numpy – data handling
• matplotlib – visualization
• OpenRouter - used to run LLM prompts and generate python scripts
• E2B sandbox - execute generated python scripts

---

⚡ Quick Start

Follow these steps to set up and run the Autonomous ML Agent locally:

1. Clone the repository

git clone https://github.com/utkarshkumar7/Autonomous-ML-Agent.git
cd Autonomous-ML-Agent

2. Create a virtual environment

python -m venv venv
source venv/bin/activate   # On macOS/Linux
venv\Scripts\activate      # On Windows

3. Install required dependencies

pip install -r requirements.txt

4. Add your OpenRouter + E2B API keys in .env file (private)

"OPENROUTER_API_KEY" = #ADD YOUR API KEY
"E2B_API_KEY" = #ADD YOUR API KEY

5. Run the application locally!

# It will run on http://localhost:8501 by default
streamlit run main.py

---

🧭 Architecture & Process Flow

Below is the end-to-end flow for the Autonomous ML Agent. It uses an LLM (e.g., DeepSeek via OpenRouter) to generate Python code, executes it inside two isolated E2B sandboxes, and then produces both a metrics dataframe and a natural-language summary with a model leaderboard.

Flow Diagram

🧭 Architecture & Process Flow

flowchart TD
  U[User / CLI]
  REG[Agent Orchestrator]
  OR[OpenRouter LLM]
  GEN[LLM Code Generation]
  SB1[E2B Sandbox 1 - Data Cleaning]
  SB2[E2B Sandbox 2 - Modeling and Evaluation]
  ART[Artifacts: cleaned.csv, model_results_df, charts]
  SUM[LLM Summarizer]
  VIZ[Leaderboard and Report]

  U -->|dataset.csv + prediction column selection on UI| REG
  REG -->|called by llm_client.py with prompts.py input| OR
  OR --> GEN
  GEN -->|script JSON unpacked as .py script| REG

  REG -->|run cleaning script| SB1
  SB1 -->|cleaned.csv + logs| ART

  REG -->|run training script with cleaned_data.csv| SB2
  SB2 -->|model results dataframe as JSON object| ART

  REG -->|model_results_df| SUM
  SUM -->|natural language summary + insights + leaderboard| VIZ

  ART --> VIZ

Loading