deepoptimizer 0.1.0rc1

AI-powered ML code analysis and optimization suggestions

DeepOptimizer: Pre-flight Checks for ML Code ✈️

Created by Willy Nilsson

Inspired by how pilots use checklists to ensure safety, DeepOptimizer provides systematic analysis of ML code before training. It combines fast pattern detection with Gemini AI insights to catch issues and suggest optimizations.

🚀 Project Components

1. CLI Tool (deepoptimizer/)

A command-line tool that analyzes Python ML code:

• Fast rule-based detection of common ML bugs
• Gemini AI-powered deep analysis
• 70+ optimization techniques in knowledge base
• Beautiful terminal output with multiple export formats

2. Web Interface (website/)

Static showcase site with:

• Project overview and installation guide
• Interactive technique demonstrations
• Examples of detected issues
• Author information

📦 Quick Start

CLI Tool (Recommended for developers)

# Install
pip install deepoptimizer  # Coming soon to PyPI

# Or install from source
git clone https://github.com/willynilsson/deepoptimizer.git
cd deepoptimizer
pip install -e .

# Set up API key
echo 'GEMINI_API_KEY=your-api-key-here' > .env

# Analyze code
deepoptimizer analyze model.py
deepoptimizer analyze ./my_project --no-llm  # Fast, rule-based only
deepoptimizer techniques --search attention   # Browse optimization techniques

Web Demo

Visit the live demo to see the optimization process in action.

Full Stack Development

# Clone repository
git clone https://github.com/willynilsson/deepoptimizer.git
cd deepoptimizer

# Backend setup
cd backend
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
python manage.py migrate
python manage.py runserver

# Frontend setup (new terminal)
cd website
npm install
npm run dev

🎯 What It Does

DeepOptimizer helps ML engineers by:

1. Finding Bugs: Detects common issues like missing model.eval(), memory leaks, wrong loss functions
2. Improving Performance: Suggests optimizations like mixed precision, better batch sizes, efficient attention
3. Teaching Best Practices: Provides research-backed suggestions with implementation examples
4. Saving Time: Automated analysis that would take hours to do manually

Example Issues Detected

# ❌ Missing model.eval() during validation
def validate(model, dataloader):
    for batch in dataloader:  # model still in training mode!
        output = model(batch)

# ❌ Memory leak from accumulating tensors
total_loss = 0
for batch in dataloader:
    loss = criterion(model(batch), target)
    total_loss += loss  # Keeps computation graph!

# ❌ Wrong loss function
loss = F.mse_loss(output, labels)  # MSE for classification!

# ✅ DeepOptimizer detects these and provides fixes

📊 Features

CLI Features

• Multi-severity analysis: Errors, warnings, and suggestions
• Project-wide analysis: Analyze entire codebases
• Export formats: Terminal, JSON, Markdown
• Configuration files: .deepoptimizer for project settings
• Offline mode: Use --no-llm for fast rule-based analysis

Detection Capabilities

• Training/validation bugs (missing eval mode, no_grad)
• Memory leaks and inefficiencies
• Wrong loss functions for tasks
• Suboptimal hyperparameters
• Missing optimizations (mixed precision, etc.)
• Architecture issues (no skip connections, etc.)

Knowledge Base

• 70+ optimization techniques
• Conflict detection between techniques
• Framework-specific implementations
• Hardware-specific optimizations
• Research paper citations

🛠️ Architecture

DeepOptimizer/
├── deepoptimizer/          # CLI tool (main package)
│   ├── analyzer.py         # Main analysis orchestrator
│   ├── llm_analyzer.py     # Gemini AI integration
│   ├── rule_detector.py    # Fast pattern matching
│   ├── knowledge_base.py   # Optimization techniques
│   └── fixtures/           # 55+ ML optimization techniques
└── website/                # Static React showcase site
    └── src/components/     # Interactive demos

The CLI tool runs completely standalone with user-provided Gemini API keys.

💡 Use Cases

1. Code Reviews: Automatically catch ML-specific issues
2. Performance Tuning: Get suggestions for faster training/inference
3. Learning: Understand optimization techniques with examples
4. CI/CD Integration: Add to pipelines for automated checks

🤝 Contributing

This is a portfolio project, but contributions are welcome! Areas of interest:

• Adding more detection rules
• Expanding the knowledge base
• Improving LLM prompts
• Adding support for more frameworks

📝 License

Apache License 2.0 - see LICENSE file. Created by Willy Nilsson.

This license provides patent protection and ensures proper attribution while remaining business-friendly and open source.

🧠 About the Author

Built by Willy Nilsson. The concept was inspired by aviation psychology - how pilots use checklists to manage complex systems safely. My background in psychology and programming led to applying these principles to ML development.

Currently seeking opportunities in AI/ML tooling development.

• Personal Site: willynilsson.com
• GitHub: @willynilsson
• LinkedIn: linkedin.com/in/willynilsson