The submission explained how LLMs tokenized text, used Transformer attention to model context, and trained billions of parameters on large corpora. It described continued improvement via fine-tuning and human feedback, producing more accurate, natural outputs.

In a Google Colab notebook I cleaned messy text with ftfy, BeautifulSoup, emoji removal, contractions expansion, regex standardization, and saved a cleaned text file. I also implemented JSON flattening for nested records to produce a structured DataFrame.

The project converted multi-page SDF PDFs into machine-readable data. I used Python in a Colab notebook to extract text, bounding boxes, and field-level values (dates, vendor, doc type, revision, signature timestamp) and output structured records.

I ran Tesseract, PaddleOCR, and EasyOCR on the same scanned supplier PDF, inspected output differences, and documented accuracy, layout preservation, setup issues, and edge cases. The deliverable concluded PaddleOCR read labels and values together best, with Tesseract catching some sideways text.

The project addressed document retrieval for technical documents. I selected a small local embedding model, chunked text into 512-token pieces with 64-token overlap, implemented hybrid vector+BM25 retrieval, added query expansion and a cross-encoder reranker, and produced ranked answers with…

I tested three RAG configurations against a Sample SDF Document, varied top_k, similarity threshold, and a reranker, and recorded retrieved chunks, best excerpts, and confidence. Configuration B removed irrelevant page noise and produced the most precise answer about QC test methods.