Developed a massively parallelized DIC algorithm using C++, CUDA, and OpenCV, improving performance and accuracy in material testing. Built automated data pipelines with Python, NumPy, and Pandas for consistent preprocessing and analysis. Performed feature engineering to enhance dataset quality for ML models. Coordinated between mechanical, software, and data teams to deliver integrated, production-ready solutions and contributed to R&D in computer vision and data automation.

Fine-tuned Falcon-7B using QLoRA on Arabic legal datasets to build a domain-specific legal reasoning model.

Fine-tuned the Falcon-7B-Instruct model using QLoRA to handle Jordanian criminal-law Q&A tasks in Arabic. Designed full data pipelines for instruction formatting, Arabic preprocessing, and dataset quality control. Used W&B for experiment tracking, hyperparameter tuning, and performance visualization. Trained on cloud GPUs (Google Colab + TensorDock) with quantization and LoRA adapters for efficient resource usage. Evaluated model performance against GPT baselines, achieving improved domain-specific accuracy and consistency.

Built a production-grade AI assistant using MCP, FastAPI, LangChain, and Qwen LLM with both voice and text interfaces.

Developed a containerized AI assistant capable of browsing and reasoning over structured public-service data through the Model Context Protocol (MCP). Designed a modular architecture where FastAPI orchestrates tool calls from a Qwen-based agent, while a local FastMCP server exposes service metadata as dynamic tools. Implemented both voice (Vosk) and text interfaces, created simulated data APIs, and built a fully interactive chatbot UI with session persistence and streaming. Authored complete architecture diagrams and documentation. The system consistently outperformed a traditional RAG chatbot by using structured tool invocation rather than static retrieval.

Rebuilt a full Digital Image Correlation algorithm from scratch using CUDA and C++, achieving industrial-grade accuracy and major speed improvements for commercial material-testing software.

Re-engineered an entire legacy DIC system by re-deriving mathematical formulations and implementing optimized CUDA kernels to achieve correctness and high performance. Integrated OpenCV preprocessing workflows for robust handling of high-resolution experimental images. Progressed rapidly from minimal C++/CUDA experience to delivering a production-ready implementation that significantly improved speed, accuracy, and robustness. The solution was adopted into Material Insights’ commercial software, enabling reliable strain-field computation for material testing workflows.