Experience

Research Assistant (Part-Time)

• Evaluated the performance of Large Language Models (LLMs) across various downstream tasks in low-resource languages like Bangla using techniques including prompting, fine-tuning, etc..
• Pretrained a new Bangla Language Model Bangla-Llama-2-7B based on meta’s Llama-2 7B model with a large Bangla corpus of around 12 GB using LORA methodology. [Code] [HF HUB]
• Currently working on training instruction-following LLaMA-based models for the Bangla language, focusing on enhancing natural language understanding and generation in a low-resource setting.

Machine Learning Engineer II

• Researching the applications of Vision-Language Models (VLMs) for activity recognition in videos to enhance remote patient monitoring in medical care facilities.

Machine Learning Engineer

• Working on a project to develop a computer vision-based product for remote patient monitoring within medical care facilities.
• Led the development of a video-redaction system that enables precise redaction of target individuals in a video using Person Detection, Instance Segmentation models, and Multi-object Trackers.
• Designed efficient stream-processing pipelines with optimized computer vision models using NVIDIA Deepstream SDK, enabling real-time video analytics on NVIDIA Jetson devices.
• Contributed to a U.S. patent for a non-invasive, real-time identification and redaction system in a monitored environment.

Associate Machine Learning Engineer

• Explored recent researches regarding various deep learning-based Computer Vision tasks such as Object Detection, Face Recognition, Pose Estimation, Segmentation, Activity Recognition etc. for different scenarios.
• Experimented with cutting-edge vision based Activity Recognition models on established benchmark datasets and assessed the real-time performance of these models on live camera feeds.
• Worked with a variety of depth sensors to utilize depth data in evaluating the effectiveness of depth-map based Human Activity Recognition models.