I am deeply passionate about Data Science and currently serve as a Data Scientist in the Genistat team at SELISE Digital Platforms, Dhaka. In this dynamic role, I actively contribute to leveraging data-driven insights to solve complex challenges, with a particular focus on media data science, computer vision, and natural language processing (NLP). This involvement allows me to apply cutting-edge techniques in machine learning and analytics. I am enthusiastic about pushing the boundaries of what data can reveal and exploring innovative solutions in the ever-evolving and dynamic field of Data Science.
• Bachelor of Science - BSc in Computer Science & Engineering, Rajshahi University of Engineering & Technology
• Higher Secondary School - Science, New Govt. Degree Collage, Rajshahi
Research Data Scientist, Genistat AG (Sep 2022 – Present)
Led the development of a cutting-edge video cropping system, utilizing advanced tiny object detection, image and signal processing techniques. Engineered the system to identify and crop videos by intelligently prioritizing relevant objects and content in each frame. Implemented dynamic aspect ratio conversion to preserve vital information and ensure superior viewing experiences across various devices. Managed the entire project from concept to execution, demonstrating exceptional technical expertise and strong project management skills.
Managed the collection of video data and conducted the training and implementation of a cutting-edge action-spotting deep learning model to extract in-game events from FIFAe game live streams.
Led the efforts in collecting and processing datasets and successfully trained a high-performance object detection model for the specific task of identifying similar category objects in live video streams.
AI Programmer, AI R&D Team (Feb 2021 – Aug 2022)
Conducted initial research on building full-scale Bengali handwritten OCR from scratch using own dataset. Led the data collection (over 250k handwritten characters), processing and character classification training phases.
Conducted research on extracting information from structured and unstructured document images using Bengali and English OCR, Amazon Textract, Text Summarizing, Bengali and English name entity recognition (using Spacy 3.0) techniques. Led the collection of unstructured document data, cleaning, and NER training processes.
Developed a data extraction system using Amazon Textract and Python for accurately retrieving relevant table data from image.
Languages: Python, C, C++, MATLAB, SQL, MySQL, OpenGL, Assembly, PHP
Libraries : OpenCV, Keras, PyTorch, TensorFlow, Pandas, Scipy, Scikit-learn, Plotly
Technologies: Heroku, Amazon Textract, Spacy, Laravel
IDE & Tools: Visual Studio Code, PhpStorm, CodeBlocks, Jupyter, Google Colab, Overleaf, ChatGPT, GitHub, Google Docs, Sheets, Drive
Version Control: Git, BitBucket
I have had the opportunity to work on a number of interesting research projects during my undergraduate. During my undergraduate thesis on computer vision and medical imaging, I was supervised by Professor Dr. Al Mehedi Hasan from the Department of Computer Science and Engineering at RUET. Additionally, I collaborated with Professor Dr. Jungpil Shin from the Department of Computer Science and Engineering at the University of Aizu, Japan. This experience significantly contributed to my research interests.
Research Interest: Deep learning, machine lerning, Medical Imaging, Computer Vision, NLP and data mining
My Publications-
Chest X-ray interpretation is very crucial to detect cardiothoracic and pulmonary abnormalities. This time-consuming and tedious task should be error-free, fast, and reliable otherwise a single mistake may cause serious harm to patients. Recently, deep learning has achieved radiologist-level performance in chest X-ray interpretation. A CAD (Computer-aided detection system) can help radiologists to review chest X-rays fast and accurately. In our research, we trained and evaluated various deep convolutional neural networks (CNN) architectures to detect potential lung opacity from chest X-rays. We observed how strongly architectures could differentiate lung opacity from normal and other abnormal chest X-rays. In these circumstances, A CNN based model (Xception) achieved 91.0% AUC along with 83.95% accuracy. Moreover, we also observed models that achieved a better performance on lung opacity vs normal chest X-ray classification (excluding abnormal class) where Xception achieved 99.1% AUC, 97.19% sensitivity, and 95.71%accuracy. Therefore, the purpose of this study is to investigate the classification ability of deep CNN architectures which helps to develop an automatic lung opacity detection system. [link]
Pneumonia is still a serious threat for children including newborns. Each year many children died of pneumonia. Physicians diagnose pneumonia through some process including reviewing chest X-rays of patients. While reviewing, a single diagnostic mistake may cause a serious threat and do significant harm to patients. In recent years, Computer-aided detection system (CAD) and medical image classification are progressively turning into another research territory. CAD can reduce the physician's effort and help to review chest X-rays fast and error-free. Currently, Researchers build various models to detect pneumonia from chest X-rays. However, there is still a lack of computationally efficient models to diagnose pediatric pneumonia. Further, some off-the-shelf or pre-trained models are not always suitable for mobile and embedded vision applications since these models are not lightweight. In our research, a lightweight convolutional neural network model was built from scratch using basic building blocks which able to learn lung texture features and detect pediatric pneumonia. Our proposed model performance was compared with some off-the-shelf models. The proposed model achieved the best AUC (99.0%), test accuracy (94.6 %), F1 (94.7 %), precision (93.2 %) and specificity (93.1%) scores. Moreover, Several data augmentation algorithms were employed to increase the model's classification ability. [link]
Being electrocardiogram already an established method for analyzing cardiac health, it gained many researchers’ interests to classify heartbeats accurately. In spite of having numerous works in this field, it still lacks obtaining high accuracy scores. In this paper, some well-known machine learning approaches are used by tuning and compared with other state-of-the-art related methodologies. The datasets used in this research work, are highly imbalanced and handled with penalizing the loss value of the Artificial Neural Network (ANN) by assigning class weights. Two different enriched ECG datasets are selected for this research. They are MIT-BIH Arrhythmia which contains five classes and PTB Diagnostic ECG which contains two classes. About 98.06% and 97.664% accuracy are achieved with proposed approaches for MIT-BIH Arrhythmia and PTB Diagnostic ECG dataset respectively. Both cases this research outperforms all the other state-of-the-art methodologies. [link]
• Convolutional Neural Networks in TensorFlow: deeplearning.ai, Coursera. [Credentials]
• The Nuts and Bolts of Machine Learning: : Google, Coursera. [Credentials]
• Crash Course on Python: Google, Coursera. [Credentials]
• AI for Medical Diagnosis: deeplearning.ai, Coursera. [Credentials]
• Database Management Essentials: University of Colorado System, Coursera. [Credentials]