Anwar Al Sheikh

Led AI initiatives including conversational AI with RAG + LLM pipelines, real-time computer vision modules for blink detection, smile analysis, and 3D head pose estimation. Designed an advanced preprocessing pipeline (perceptual hashing, cosine similarity, Annoy) for a 300K+ image dataset. Implemented synthetic data generation to rebalance a 3-class vehicle color dataset (Yellow, Orange, Grey), achieving 95% accuracy. Deployed models with quantization and converted PyTorch models to ONNX/TRT for performance. Built model clients and API wrappers for production-ready services and integrated CI/CD workflows in Azure DevOps.

Developed AI-powered chatbots for question answering, summarization, and legal clause extraction using RAG + LLM pipelines to improve response accuracy and contextual understanding. Integrated real-time facial expression detection and face-action detection (blink detection via Eye Aspect Ratio, smile detection from facial landmarks, and 3D head pose estimation). Designed and optimized data preprocessing with perceptual hashing for duplicate detection and Cosine Similarity/Annoy-based filtering to reduce redundancy in a 300,000-image dataset. Applied segmentation to remove backgrounds and used HSV-based synthetic data generation to augment vehicle color classification. Led large-scale data acquisition and annotation with CVAT. Implemented model quantization and deployment (PyTorch to ONNX and TensorRT), and exposed models via API wrappers. Version-controlled work with Git/GitHub and various Linux operations.

Led end-to-end AI initiatives including conversational AI using RAG + LLM pipelines and real-time computer vision for action recognition (blink, smile, head tracking) and attendance. Built model clients and production-ready API wrappers; implemented CI/CD pipelines in Azure DevOps for Linux-based deployments. Conducted thorough data preprocessing and optimization, including a duplicate image filtration system for 300K+ images and synthetic data generation to rebalance a 3-class vehicle color dataset, achieving 95% validation accuracy. Spearheaded model deployment and quantization to accelerate inference via ONNX/TRT. Coordinated data acquisition and annotation workflows using CVAT.