The submission summarized how LLMs predicted tokens using numerical representations, described Transformer self-attention, and outlined pretraining, supervised fine-tuning, and reinforcement learning from human feedback to improve responses.

The work showed Python notebooks in Google Colab that cleaned tabular data with Pandas, normalized nested JSON into flat tables, performed text cleaning and normalization, and applied image preprocessing (denoising, thresholding, contrast) to improve OCR input.

The project processed SDF PDFs with PyMuPDF, extracted every word with bounding boxes, and then used keyword anchors and proximity logic to locate fields. The submission produced JSON word output and listed extracted fields including manufacture and expiration dates.

I ran Tesseract, PaddleOCR, and EasyOCR on a scanned supplier PDF, recorded text-region counts, per-engine confidence scores, and detected key pharmaceutical terms. I reported that PaddleOCR had the highest accuracy and confidence, with metrics and a comparison table.