Project 1: How AI Reads and Understands Mortgage Documents

I examined how AI comprehends mortgage documents by studying Machine Learning, Deep Learning, LLMs, and NLP. I also investigated Computer Vision and OCR techniques for extracting information from scanned documents, addressing challenges in the mortgage industry and emphasizing the need for accurate

Project 3: Data Extraction from Documents Using Python

Before AI can make sense of documents, it needs clean, structured data. In this project, you’ll learn how to turn both digital and scanned PDFs into machine-readable formats using Python. You’ll explore tools like PyMuPDF and pdfplumber to extract text from multi-page mortgage files. Then, you’ll apply field heuristics—like regex patterns, anchor phrases, and layout clues—to identify key fields such as borrower name, salary, and loan amount.

Project 5: Introduction to Retrieval-Augmented Generation (RAG)

AI models can process vast amounts of text, but finding the right information quickly is a challenge. In this project, you’ll learn how to build a Retrieval-Augmented Generation (RAG) pipeline using LlamaIndex to help AI retrieve relevant data from large document sets.

Project 7: Blob Processing, Classification, and Routing

Mortgage files often come as massive, unstructured "blobs" that include dozens of document types in a single PDF. In this project, you’ll design a system to split unstructured blobs into individual documents, classify each one, and route them to the appropriate extraction logic. You’ll use layout-based clues and rule-based methods—or even simple machine learning—to separate and identify documents like pay slips, IDs, and contracts. Then, you’ll build a routing function to apply the right extraction pipeline for each document type. By the end, you’ll have created a modular, intelligent system that mirrors real-world document workflows in mortgage automation.

Project 9: Final Integration, Testing, and Evaluation

It’s time to bring everything together. In this final project, you’ll integrate all components into a complete, end-to-end document intelligence system. You’ll run your pipeline on a full mortgage blob, from preprocessing and OCR to extraction, RAG retrieval, and user interaction. Then, you’ll rigorously test the system’s performance, evaluating accuracy, speed, and reliability across different document types. Finally, you’ll package your work into a clear, professional demo—including a Colab repo, documentation, and walkthrough video—that showcases your system’s real-world potential. By the end, you’ll have a fully functional AI-powered document automation platform ready to present to peers, mentors, or even hiring managers.

Project 2: Learn to Work with Python for AI-Powered Document Processing

I utilized Python for data processing, focusing on cleaning and organizing mortgage documents to prepare them for AI automation. I explored Google Colab, learned Python basics, and implemented data preparation techniques that enhanced the accuracy of Optical Character Recognition (OCR) for mortgage

Project 4: Advanced OCR Comparison and Layout-Aware Extraction

OCR (Optical Character Recognition) is at the heart of document intelligence. In this project, you’ll get hands-on with three OCR engines— Tesseract, PaddleOCR, and EasyOCR —to see how they perform on complex mortgage documents. You’ll start by learning how to clean and extract text from scanned PDFs using Tesseract. Then, you’ll explore layout-aware tools that better preserve formatting and structure. Finally, you’ll run comparisons and recommend the most effective tool for real-world automation. By the end, you’ll be able to confidently select and justify OCR tools based on both technical and business needs.

Project 6: Advanced RAG and Open-Source Experiments

In this project, you’ll take your RAG pipeline to the next level by improving precision and experimenting with open-source LLMs. First, you’ll explore advanced chunking techniques—like overlapping and tuning—to improve document segmentation. Then, you’ll implement metadata-based filtering so your system retrieves more contextually relevant answers. You’ll also evaluate the performance of open-source models like Mistral and Phi-2 compared to proprietary options. By the end, you’ll have a fine-tuned, high-performing RAG pipeline optimized for real-world mortgage documents and the ability to justify your design decisions based on retrieval accuracy and model efficiency.

Project 8: Gradio Chatbot with RAG Integration

I created an interactive chatbot integrated with my RAG pipeline using Gradio. This allowed users to query documents through a web-based interface, modularizing the pipeline and facilitating easy integration of various language models for enhanced user interaction.