Ubaid ur Rehman

Automating structured data extraction from insurance forms (NAF, CNIC, Proposals) was bottlenecked by high latency and complex nested schemas.

Fine-tuned GLM-OCR using LoRA (fp16, rank 16) on a 636-sample dataset; deployed via SGLang on Vast.ai GPUs; integrated with Azure Document AI, Django multi-threaded PDF rendering, and a React key-matcher with alias mapping and recursive search.

1.2 seconds/page inference latency in enterprise production with recursive field search across nested form schemas.

Python, PyTorch, GLM-OCR (1.1B), LoRA, Azure Document AI, SGLang, Django, React, Vast.ai

Skewed form layouts and complex tables caused data entry bottlenecks, requiring precise oriented bounding-box detection and semantic field extraction.

Designed and implemented a multi-agent pipeline combining YOLOv11 Oriented Bounding Boxes (OBB), Tesseract OCR, and LLM integration. Built full-stack Django + React with JWT auth, RBAC, and ONNX Runtime.

69% field-detection accuracy at 10–15s end-to-end latency per document with ONNX Runtime for optimized inference.

Python, YOLOv11 OBB, Tesseract OCR, LLMs, Django, React, JWT, RBAC, ONNX Runtime, PostgreSQL

Manual job searching across fragmented listings consumed 15+ hours weekly with delayed alert notifications.

Built an end-to-end automated scraper using Python and Selenium orchestrated with n8n workflows for scheduled extraction and email digest dispatch.

Delivers 500+ structured LinkedIn job listings per week via automated email digest, cutting manual job-search effort by 90%.

Python, Selenium, n8n, SMTP, Git, Linux (Ubuntu)

Property buyers and investors lacked transparent, accessible valuation models based on localized real estate data.

Trained 5 ML regression models on a self-scraped dataset, engineered custom features, and ensembled predictions via soft-voting.

Achieved 89–92% prediction accuracy; deployed as a public web application with Flask and Streamlit.

Python, Scikit-learn, XGBoost, Pandas, NumPy, Flask, Streamlit