TuneLab

Fully Autonomous ML Engineer Agent

A production-ready, fully autonomous machine learning pipeline for tabular datasets. Build high-quality models end-to-end without human intervention.

Features

• 🔄 Fully Autonomous: Zero human intervention required
• 💾 Strategy Memory: Learns from past runs via dataset fingerprinting
• 📊 Complete Pipeline: From raw data to production model
• 📝 Professional Docs: Auto-generated Markdown reports
• 🖥️ CPU-First: Optimized for CPU, GTX 1650 friendly (no GPU required)
• 🆓 100% Open Source: No proprietary dependencies

---

🚀 Quick Start

Installation

# Clone or download this repository
git clone <your-repo-url>
cd ml-agent

# Install dependencies
pip install -r requirements.txt

Basic Usage

# Run on your dataset
python ml_agent.py path/to/your/data.csv

That's it! The agent will:

1. Analyze your data
2. Engineer features
3. Train multiple models
4. Optimize hyperparameters
5. Generate reports and plots
6. Save the best model

Example with Sample Data

# Generate sample data first
python example_usage.py --generate-data

# Run the agent
python ml_agent.py sample_data/iris.csv

---

📁 Output Structure

After running, you'll find everything in the outputs/ directory:

outputs/
├── models/
│   └── final_model.joblib          # Trained model ready for production
├── plots/
│   ├── feature_importance.png      # Top features visualization
│   └── metric_comparison.png       # Model performance comparison
├── reports/
│   ├── overview.md                 # Project summary
│   ├── data_analysis.md            # Data insights
│   ├── modeling.md                 # Model selection details
│   └── results.md                  # Final results & recommendations
└── strategy/
    └── <fingerprint>.json          # Reusable strategy for this dataset

---

🎯 How It Works

1. Dataset Fingerprinting

Generates a unique fingerprint based on:

• Dataset shape
• Column names and types
• Missing value patterns

If you've run the agent on similar data before, it loads the successful strategy from memory.

2. Data Understanding

Automatically detects:

• Target column (last column by default)
• Problem type (classification vs regression)
• Missing values
• Feature types (numerical vs categorical)
• Class imbalance

3. Feature Engineering

Applies robust preprocessing:

• Numerical: Median imputation
• Categorical: Most frequent imputation + label encoding
• No leakage: All transformations fit only on training data

4. Model Selection

Trains multiple baseline models:

Classification:

• Logistic Regression
• Random Forest
• Extra Trees
• Gradient Boosting

Regression:

• Linear Regression
• Ridge Regression
• Random Forest
• Extra Trees
• Gradient Boosting

5. Hyperparameter Optimization

Uses Optuna for Bayesian optimization:

• 30 trials
• 3-fold cross-validation
• Automatic parameter search

6. Artifacts & Documentation

Generates:

• Saved models (.joblib)
• Visualizations (.png)
• Professional Markdown reports
• Reusable strategy files

---

💻 Advanced Usage

Specify Target Column

from ml_agent import MLAgent

agent = MLAgent(
    data_path="data.csv",
    target_col="target",  # Specify target column
    problem_type="classification"  # Or "regression"
)

agent.run()

Customize Parameters

agent = MLAgent(
    data_path="data.csv",
    output_dir="my_outputs",
    max_iterations=5,
    target_metric_threshold=0.95,
    improvement_threshold=0.01
)

agent.run()

Load and Use Trained Model

import joblib
import pandas as pd

# Load model package
model_pkg = joblib.load('outputs/models/final_model.joblib')

model = model_pkg['model']
preprocessor = model_pkg['preprocessor']
feature_names = model_pkg['feature_names']

# Load new data
new_data = pd.read_csv('new_data.csv')

# Make predictions
predictions = model.predict(new_data)

---

🛠️ Tech Stack

Core (Required)

• Python 3.10+
• pandas
• numpy
• scikit-learn
• matplotlib
• joblib

Optional (Recommended)

• seaborn (better plots)
• optuna (hyperparameter tuning)

Constraints

• ✅ 100% open source
• ✅ CPU-optimized
• ✅ No GPU required
• ✅ GTX 1650 compatible if GPU available
• ❌ No proprietary software
• ❌ No paid APIs

---

📊 Supported Models

Classification

Model	Speed	Accuracy	Interpretability
Logistic Regression	⚡⚡⚡	⭐⭐	⭐⭐⭐
Random Forest	⚡⚡	⭐⭐⭐	⭐⭐
Extra Trees	⚡⚡	⭐⭐⭐	⭐⭐
Gradient Boosting	⚡	⭐⭐⭐	⭐

Regression

Model	Speed	Accuracy	Interpretability
Linear Regression	⚡⚡⚡	⭐⭐	⭐⭐⭐
Ridge Regression	⚡⚡⚡	⭐⭐	⭐⭐⭐
Random Forest	⚡⚡	⭐⭐⭐	⭐⭐
Extra Trees	⚡⚡	⭐⭐⭐	⭐⭐
Gradient Boosting	⚡	⭐⭐⭐	⭐

---

📖 Example Workflow

1. Prepare Your Data

Your CSV should have:

• Features in columns
• Target variable (typically last column)
• Header row with column names

feature1,feature2,feature3,target
1.2,3.4,cat,0
2.3,4.5,dog,1
...

2. Run the Agent

python ml_agent.py my_data.csv

Output:

🤖 ML Agent Initialized
📁 Output Directory: outputs

📊 Loading dataset: my_data.csv
   Shape: (1000, 5)
   Columns: ['feature1', 'feature2', 'feature3', 'target']
🔑 Dataset Fingerprint: a3f5d8c9b2e1f4a7

🔍 Analyzing dataset...
   Auto-detected target: target
   Auto-detected problem type: classification
   ✅ No missing values
   Numerical features: 2
   Categorical features: 1

🔧 Engineering features...
   Processing 2 numerical + 1 categorical features
   ✅ Train: 800 samples
   ✅ Test: 200 samples

🎯 Training baseline models...
   Training Logistic Regression... accuracy=0.8500
   Training Random Forest... accuracy=0.9200
   Training Extra Trees... accuracy=0.9100
   Training Gradient Boosting... accuracy=0.9350

   🏆 Best model: Gradient Boosting (accuracy=0.9350)

⚙️  Optimizing hyperparameters...
   Best trial score: 0.9425
   ✅ Optimized model score: 0.9450

💾 Model saved: outputs/models/final_model.joblib

📊 Generating plots...
   ✅ Feature importance plot saved
   ✅ Model comparison plot saved

📝 Generating reports...
   ✅ All reports generated

💾 Strategy saved to memory: a3f5d8c9

============================================================
✅ PIPELINE COMPLETE
============================================================

📁 All outputs saved to: outputs/
🏆 Final model score: 0.9450
💾 Model: outputs/models/final_model.joblib
📊 Reports: outputs/reports/
📈 Plots: outputs/plots/

3. Review the Outputs

Check the generated reports:

• outputs/reports/overview.md - Quick summary
• outputs/reports/data_analysis.md - Data insights
• outputs/reports/modeling.md - Model details
• outputs/reports/results.md - Final results

View visualizations:

• outputs/plots/feature_importance.png
• outputs/plots/metric_comparison.png

4. Use the Model

import joblib

# Load and use
model_pkg = joblib.load('outputs/models/final_model.joblib')
predictions = model_pkg['model'].predict(new_data)

---

🎓 Design Philosophy

1. Simplicity Over Complexity

• Use simple, explainable models first
• Avoid overfitting with regularization
• Prefer interpretability when possible

2. CPU-First Architecture

• No GPU required (though compatible)
• Optimized for standard hardware
• Works on laptops and servers alike

3. Full Transparency

• Every decision is documented
• Complete audit trail in reports
• Reproducible with fixed random seeds

4. Production Ready

• Save everything needed for deployment
• Include preprocessing in model package
• Professional documentation for handover

5. Autonomous Operation

• Zero human intervention
• Automatic problem detection
• Self-documenting workflows

---

🔧 Customization

Add Custom Models

Edit ml_agent.py in the train_models() method:

if self.problem_type == 'classification':
    models = {
        'Logistic Regression': LogisticRegression(...),
        'Random Forest': RandomForestClassifier(...),
        # Add your model here:
        'SVM': SVC(...),
    }

Change Metrics

Modify the metric calculation in train_models():

# For classification
metrics = {
    'accuracy': accuracy_score(self.y_test, y_pred),
    'f1': f1_score(self.y_test, y_pred, average='weighted'),
    # Add custom metrics
}

Adjust Hyperparameter Search

Modify optimize_hyperparameters():

# Change number of trials
study.optimize(objective, n_trials=50)  # Default: 30

# Adjust cross-validation folds
scores = cross_val_score(..., cv=5)  # Default: 3

---

📝 Report Examples

Overview Report

• Quick project summary
• Tech stack used
• Best model and score
• File structure

Data Analysis Report

• Dataset statistics
• Missing value analysis
• Feature type breakdown
• Target distribution

Modeling Report

• All models tried
• Performance comparison
• Selection rationale
• Hyperparameter tuning results

Results Report

• Final metrics
• Usage instructions
• Next steps recommendations
• Reproducibility guide

---

🐛 Troubleshooting

"Optuna not available"

pip install optuna

Or continue without it (hyperparameter tuning will be skipped).

"Memory Error"

For large datasets, reduce n_estimators in models:

'Random Forest': RandomForestClassifier(n_estimators=50)  # Default: 100

"FileNotFoundError"

Make sure your CSV path is correct:

python ml_agent.py /full/path/to/data.csv

---

🤝 Contributing

This is a fully autonomous agent - improvements welcome!

Areas for enhancement:

• Additional model types
• Advanced feature engineering
• Custom metric support
• Multi-class calibration
• Time series support

---

📄 License

Open source - use freely for any purpose.

---

🙏 Acknowledgments

Built with:

• scikit-learn - Amazing ML library
• Optuna - Hyperparameter optimization
• pandas - Data manipulation
• matplotlib - Visualization

---

📞 Support

For issues or questions you can contact me at pranavbansode2604@gmail.com

1. Check the generated reports in outputs/reports/
2. Review the troubleshooting section
3. Examine the code comments in ml_agent.py

---

Built for production. Designed for autonomy. Optimized for simplicity.

🤖 Let the agent do the work.