The Legal & Compliance Checker

Auto-Disclaimer Prototype
Brand Safety Compliance enforcer
Profanity Filter

Inspiration

In today's digital age, content creators face mounting pressure to ensure their work is compliant, inclusive, and legally sound across global markets. A single piece of offensive content or copyright violation can damage brand reputation, trigger legal action, or alienate audiences. We noticed that designers and marketers often lack real-time tools to catch these issues before publication.

Pixel-Proof was born from a simple question: What if AI could act as a guardian angel for content creators, catching compliance risks before they become costly mistakes?

We envisioned an Adobe Express add-on that would seamlessly integrate into creative workflows, providing instant feedback on hate speech, copyright violations, and brand safety—all powered by cutting-edge machine learning.

What It Does

Pixel-Proof is an enterprise-grade Adobe Express add-on that provides real-time content compliance analysis through three core capabilities:

🛡️ Hate Speech Detection
- Analyzes text using fine-tuned HateBERT transformer models
- Provides segment-level analysis with confidence scores
- Visually highlights problematic content in red directly within documents
- Processes both text nodes and OCR-extracted text from images
🔍 Copyright & Trademark Detection
- Recognizes logos using color descriptor-based matching
- Searches against a database of 4.2 million brand logos
- Detects trademark violations and copyrighted brand usage
⚖️ Legal Compliance Tools
- Auto-generates multilingual legal disclaimers (English, Spanish, French, German)
- Customizable disclaimer templates for different industries
- Brand guideline compliance checking

How We Built It

Architecture: We designed a three-tier architecture optimized for performance and scalability:

Adobe Express Add-on (React + TypeScript)
         ↓
Node.js Backend (Express Server + MCP Server)
         ↓
Python ML Service (Flask + HateBERT)

Challenges We Faced

1. Adobe Express SDK Limitations

Challenge: The SDK had limited documentation for advanced text manipulation
Solution: Extensive experimentation with the Document API, reverse-engineering examples, and creative workarounds for text styling

2. ML Model Performance

Challenge: HateBERT model (110M parameters) took 5-10 seconds to load on startup
Solution: Implemented model pre-loading on service startup and kept the model in memory for fast inference

3. Cross-Origin Communication

Challenge: Sandbox isolation prevented direct API calls from the document context
Solution: Designed a runtime API proxy pattern for secure UI-to-sandbox communication

4. Logo Recognition Accuracy

Challenge: Initial color-based matching had ~20% false positive rate
Solution: Refined color descriptor algorithm, implemented stricter thresholds, and added chi-squared distance metrics

5. OCR Integration

Challenge: Tesseract.js processing was slow (3-5s per image)
Solution: Optimized image preprocessing, implemented parallel processing for multiple images, and added progress indicators

6. Real-time Visual Feedback

Challenge: Highlighting hate speech in documents without disrupting layout
Solution: Developed custom text node traversal algorithm that preserves document structure while applying styling

What We Learned

Technical Learnings:

Deep understanding of transformer-based NLP models and their practical deployment
Mastery of Adobe Express SDK and add-on development patterns
Advanced image processing techniques with OpenCV
Microservices architecture design for ML applications
Real-time document manipulation in constrained environments

Product Learnings: