News Impact Trainer - Emotion Classification ML Pipeline

ML Pipeline (Python + Node.js)

About

Full-stack ML application for emotion classification of news articles. Combines a Node.js pipeline fetching from 25+ news sources, a browser-based labeling UI, Claude-assisted batch labeling, and a fine-tuned DistilBERT model for 7-way emotion detection.

My Role

Built the complete pipeline: RSS feed aggregation from 25+ sources, interactive labeling interface, Claude AI integration for batch labeling, and DistilBERT model fine-tuning with ONNX export.

Core Capabilities

• RSS feed aggregation from 25+ Pakistani, international, and business news sources
• Interactive browser-based labeling UI with keyboard shortcuts and progress tracking
• Claude AI batch labeling integration for rapid dataset creation
• DistilBERT fine-tuning for 7-way emotion classification (happy, angry, anxious, sad, neutral, hopeful, surprised)
• ONNX Runtime export for 10x faster production inference
• Data quality pipeline: deduplication, junk filtering, HTML cleaning

Architecture

• Node.js RSS scraper with rss-parser for 25+ feeds
• HTML5 + vanilla JS labeling interface with localStorage persistence
• HuggingFace Transformers for DistilBERT fine-tuning
• PyTorch training loop with stratified train/val split
• ONNX export for production deployment

Tech Stack