I built a full Gradio-based RAG chatbot that ingested pharmaceutical PDFs, extracted text with PyMuPDF/Tesseract, indexed chunks in FAISS, and returned LLM answers with source citations.

I explained how large language models work and how they generate text, training tokens, transformers, attention, fine-tuning, and the risk of hallucinations.

I prepared and cleaned pharmaceutical documents with Python in Google Colab, applying data cleaning, JSON normalization, text standardization, and image preprocessing to improve OCR readiness.

I extracted text and bounding-box coordinates from SDF PDFs with PyMuPDF and used heuristics to locate date fields and other key values.

I compared Tesseract, EasyOCR, and PaddleOCR on a scanned pharmaceutical certificate and documented which engine best preserved fields, layout, and reading order.

I built a RAG pipeline evaluation comparing two open-source LLMs on a sample document, running retrieval, generation, and measured response quality and speed.

I ran and compared three RAG retriever configurations on an SDF document, logging retrieved chunks, generated answers, and per-config observations.

I built a page-level segmentation and metadata extractor that split bundled PDFs into distinct documents and tagged pages with doc_type and page_in_doc.