Developed a cutting-edge hierarchical Transformer-based model for intraoperative abdominal organ segmentation, setting a new benchmark for accuracy on the Dresden Surgical Anatomy Dataset. By designing a multi-scale cross-attention decoder with a DINOv2 backbone and a hybrid loss function, I boosted Dice scores by up to 27% on challenging organ classes, such as the pancreas and colon. Additionally, I implemented an advanced ROI extraction strategy to enhance the segmentation of smaller and partially cropped organs, significantly improving the model's performance on complex, structurally intricate regions.

Researching on a modified Diffusion-based Near-Infrared (NIR) model for colorizing night vision footage, incorporating ILVR-style refinement for improved output. The model was tested on the NIR2RGB VCIP dataset, where it achieved strong results, including a Peak Signal-to-Noise Ratio (PSNR) of 22.7 dB and a Structural Similarity Index (SSIM) of 0.67. These results demonstrated significant perceptual improvements in color fidelity and detail reconstruction, outperforming baseline GAN and U-Net approaches in terms of both accuracy and visual quality.

A GraphRAG pipeline for citation-grounded response generation over a legal knowledge graph, featuring local inference optimization.

A scalable vocabulary platform featuring AI-powered voice assistants and microservice architecture.

A scalable micro-blogging platform with real-time notifications and stateless JWT authentication.

Master the fundamentals of containerization and orchestration by learning foundational virtual environments, Docker container management, and Kubernetes basics for scalable and resilient application deployment.

Achieved 39th rank in the Machine Learning Olympiad 2024 organized by the Google Developer Group (GDG) Bangladesh, demonstrated strong proficiency in machine learning concepts, algorithms, and practical applications through a series of challenging problems and projects.

Build and train machine learning models using NumPy and scikit-learn for prediction and binary classification tasks with linear and logistic regression. Develop neural networks in TensorFlow for multi-class classification and implement decision trees and ensemble methods. Apply best practices for ML development and leverage unsupervised learning for clustering and anomaly detection. Create recommender systems using collaborative filtering and content-based deep learning, and develop a deep reinforcement learning model.

Finalist in the IEEE CS Spectrum: Code Crash 2023 competition, competiting with 9 finalists from around 100 teams, showcasing strong problem-solving abilities in data structures and algorithms in a measured time frame.