Abstract: Efficient sorting of plastic waste remains a critical bottleneck in recycling systems, with current approaches relying on manual labor or semi-automated solutions that contribute to large amounts of plastics ending up in landfills. Despite the rapid growth of the global plastic recycling market, projected to reach \$120 billion by 2030, existing sorting technologies struggle to meet demands for accuracy and throughput \cite{marketsandmarkets_recycled_plastic}. While recent ML breakthroughs show promise in waste sorting, a complete industrial-scale pipeline has been overlooked. We propose a novel, low-cost machine learning system that addresses real-world challenges in plastic sorting: varying material types, inconsistent lighting conditions, and contaminated surfaces. Our key contributions include: (1) a scalable deep learning architecture featuring two adaptive pipelines - one for data collection and another for classification, optimized for industrial deployment, (2) curation of the world's first comprehensive industrial dataset of 40,000 plastic samples, and (3) an interpretable approach leveraging Grad-CAM and t-SNE visualizations to tackle challenging cases like dark and distorted plastics. The proposed sorting system demonstrates commercial viability by processing 200 samples per hour across five plastic types common in municipal solid waste (MSW), with potential earnings of \$30 per ton.