Built an automated invoice extraction pipeline for a logistics company processing multiple scanned PDFs daily. I wrote a custom OpenCV preprocessing module handling deskewing via Hough Transform, adaptive thresholding, and contour-based border cropping. On top of that, I trained a spaCy NER model from scratch on annotated invoices with custom entities vendor, dates, line items, totals. A validation layer cross-checked extracted totals and flagged anomalies for human review. Result: 91% accuracy, 80% less manual work. I also developed the full web interface using Laravel and React.