I summarized LLM internals: tokenization to integer IDs, embeddings as high-dimensional vectors, transformer attention with multiple heads, repeated layers, and gradient-descent training. The submission produced a concise, technical explanation of how LLMs learn and predict text.

The module provided a Google Colab notebook that showed Python setup, basic data structures, control flow, functions, and an input-to-output example. I executed code cells and included runnable examples for data cleaning and image preprocessing.

The module involved extracting words and bounding boxes from SDF PDFs with PyMuPDF, reporting which pages and layouts extracted cleanly, noting dropped elements (logos, addresses), coding a separate bounding-box extractor, and recommending regex, label proximity, and bbox rules to locate date…

The assignment compared three OCR engines on a scanned pharmaceutical supplier PDF. I ran Tesseract, PaddleOCR and EasyOCR, inspected outputs and found PaddleOCR preserved layout and bounding boxes best, reporting its higher confidence and setup trade-offs.