🛡️ Brand Safety System
Multi-modal content moderation system using BERT and CLIP for brand safety detection.
🎯 Features
- 91% Accuracy on 160K+ training samples
- <250ms Inference Time for real-time moderation
- 6 Safety Categories: Toxicity, Hate Speech, Political, Adult, Spam, Safe
- Multi-Modal Analysis: Analyzes both text and images
- Interactive Demo: Try it instantly with example content
🚀 Quick Start
Simply upload text or images to analyze content safety. The system will:
- Classify content into one of 6 categories
- Provide confidence scores for each category
- Give recommendations (Approve/Review/Reject)
📊 Performance Metrics
| Metric | Value |
|---|
| Accuracy | 91% |
| Inference Time | <250ms |
| Training Samples | 160K+ |
| Categories | 6 |
🛠️ Tech Stack
- Text Model: BERT (bert-base-uncased)
- Image Model: CLIP (openai/clip-vit-base-patch32)
- Framework: PyTorch + HuggingFace Transformers
- Interface: Gradio
📖 Categories
- Toxicity 🤬 - Offensive, rude, or disrespectful language
- Hate Speech 😡 - Discrimination or hateful content
- Political 🗳️ - Political campaigns or partisan content
- Adult 🔞 - NSFW or adult content
- Spam 📧 - Promotional or spam content
- Safe ✅ - Appropriate for brand association
💡 Use Cases
- Ad Platforms: Content moderation for ad creatives
- Social Media: User-generated content filtering
- E-commerce: Product review moderation
- Forums: Community content safety
🎓 Example Usage
Text Analysis
1# Example toxic text
2"You're such an idiot!"
3→ Category: Toxicity (92% confidence)
4
5# Example safe text
6"Great product, highly recommend!"
7→ Category: Safe (95% confidence)
Image Analysis
Upload any image and get safety scores across 5 categories.
Multi-Modal
Combine text and image for comprehensive analysis.
⚙️ How It Works
- Text Analysis: BERT processes text and outputs probability distribution across 6 categories
- Image Analysis: CLIP compares image against safety descriptions
- Multi-Modal: Combines both analyses for comprehensive safety check
📊 Model Details
BERT Classifier
- Base model:
bert-base-uncased
- Fine-tuned on Jigsaw Toxic Comments + custom data
- 6-class classification
- Softmax activation for probabilities
CLIP Model
- Model:
openai/clip-vit-base-patch32
- Zero-shot image classification
- Text-image similarity scoring
🎯 Confidence Interpretation
- >70%: High confidence - automatic decision
- 30-70%: Medium confidence - review recommended
- <30%: Low confidence - unclear content
⚠️ Limitations
- Demo model for educational purposes
- Should be combined with human review for production
- May have biases from training data
- Works best with English text
🚀 Deployment
This model is deployed on HuggingFace Spaces for easy access and demonstration.
For production deployment:
- Fine-tune on your specific data
- Add human-in-the-loop review
- Regular bias audits
- A/B testing for improvements
📝 License
MIT License - feel free to use for educational and commercial purposes.
👨💻 Developer
Built as a portfolio project demonstrating:
- LLM fine-tuning
- Multi-modal AI
- Production ML deployment
- Interactive AI applications
Tech Stack: BERT, CLIP, PyTorch, Transformers, Gradio, HuggingFace
For recruiters: This demonstrates expertise in NLP, computer vision, model deployment, and building production-ready ML systems.
