Mahmudur R.

With a strong background in machine learning, computer vision, and natural language processing, I have consistently delivered real-world AI solutions across diverse industries. My professional journey spans multiple roles where I have developed and deployed intelligent systems for image and language understanding, focusing on accuracy, performance, and scalability. Currently working as a freelancer, I have built intelligent data extraction pipelines for financial applications by combining advanced computer vision and NLP methods to extract information from bank cheques with high precision. Additionally, I worked on a complex chatbot project addressing the limitations of large language models (LLMs), such as context window constraints. I implemented optimization strategies like history truncation, prompt summarization, and prompt caching, which significantly improved coherence, processing speed, and memory efficiency in generated responses. As a Data Science Fellow at Fellowship.AI, I developed a robust evaluation pipeline to measure confidence levels in LLM outputs using datasets like LiveBench and MMLU. I applied strong evaluation metrics such as Expected Calibration Error (ECE) and BERTScore and utilized stronger LLMs for benchmarking, providing valuable insights into model calibration and reliability. Previously, as a Computer Vision Engineer at Hello Llama, I played a pivotal role in building IoT-enabled safety solutions. I developed end-to-end ML pipelines and Android applications integrating BLE and real-time video streaming, designed to work with radar warnings and dashcam systems. My work focused on real-time object detection of urban infrastructure and safety violations, deploying models on NVIDIA Jetson Nano for low-latency, edge-based inference. I also created a custom helmet detection system with facial focus and chin-strap detection and used sensor mat data to recognize multiple riders on scooters. My responsibilities included complete development cycles—from data collection to annotation, model training, validation on unseen data, and deployment—with comprehensive testing to ensure robustness in production environments. Earlier in my career as a Machine Learning Engineer at Expert Consortium Ltd., I worked on automatic face recognition and liveness detection systems. I designed pipelines capable of distinguishing live individuals from photographs and automatically created datasets for unknown individuals by organizing webcam feeds, labeling them, and training recognition models using LBP. I also built a driver activity recognition system for behavioral monitoring and an object tracking tool that captured snapshots when specific spatial triggers were activated. I utilized GPU acceleration and RabbitMQ to ensure high performance and seamless message passing during training and inference. Throughout my experience, I have demonstrated a unique ability to bridge the gap between computer vision and natural language processing. I have built robust pipelines, optimized LLM performance, engineered edge-deployable vision systems, and created intelligent, real-time applications in safety, finance, and mobility. My strengths lie in developing production-ready ML systems, optimizing performance in resource-constrained environments, and delivering intelligent solutions that scale effectively and meet real-world demands. I have direct Experience working on following topics: C++ Python Scikit-learn Tensorflow Keras OpenCv Embedded device Aws Azure Flutter