Experience

Research Assistant

• Enhance Robustness of Diffusion-based Models (e.g. Inpainting) via Self-Supervised Transformers (Preparing for ECCV'24).

Undergraduate Thesis

• Designed a semi-supervised transformer enhancing semantic segmentation and depth estimation tasks through symbiosis and a dual compensating augmentation, ensuring object integrity, diversity, and controlled generation [WACV'24].
• Proposed local-global window transformers to mitigate trade-off between detailed & consistent depth map [IEEE JSEN'23].
• Introduced a dataset comprising 2.5 Million real vs. fake images (25 generators) and a novel multi-class strategy to detect both seen and unseen generators amid social-media impairments (compress and downsample) [ICIP'23].

Research Collaboration

• Developed a semi-supervised ensemble technique to identify the algorithms responsible for generating synthetic speeches amidst perturbations, using mel-spectrogram features. [In Review at ICASSP'24].

Research Collaboration

• Proposed horizontal network expansion to reduce cost while maintaining good performance in COVID-19 lesion segmentation from CT scans; all codes, diagrams, and experiments implemented [IEEE TAI]
• Published Paper on IEEE Transactions on Artificial Intelligence (Q1)
• Paper - IEEE Xplore

Undergraduate Research Assistant

• Introduced novel network for Melanoma diagnosis by concurrent multi-image comparison with patient-level contextual data.
• Proposed an efficient AF detection scheme to integrate raw ECG and discrete wavelet transformed features.
• Contributed by introducing stacking ensemble strategy, and implementing all codes, diagrams, experiments in "CovXNet" project.

Contractor

• Develop code examples and tutorials for Kaggle competitions using Keras, focusing on Computer Vision and NLP tasks.

Developer Expert

• Testing WandB in ML projects, troubleshooting issues, optimization, and creating tutorials for effective tracking workflow.
• Integrated WandB pipeline into YOLOv5 and tested it for Underwater Object Detection. (Code - wandb)

Collaboration with NVIDIA

• Worked on COVID-19 Detection and Localization from Chest X-Ray.
• Introduced a Semi-supervised Multi-Stage Transfer Learning scheme to detect abnormalities from limited data.
• Developed a fast BBox-Filter method to remove abnormal bounding boxes using statistical distribution.
• Won Best Student Team Award and ranked 4th globally in SIIM-FISABIO-RSNA COVID-19 Detection Competition.
• Code - GitHub | Video - YouTube

Kaggle Grandmaster

• Competed with professionals from leading companies like NVIDIA, HuggingFace, H20.ai, and more.
• Ranked 7th (best 5th) among 300,000 contenders and 1st in Bangladesh at the Code section.
• Won 36 Gold, 11 Silver medals in the Code section, and 1 Gold, 3 Silver medals in the Competition section.

MLDL-I - Machine Learning & Deep Learning I (Course + Lab)

• Theory and hands-on coding for Machine Learning & Deep Learning.
• ML - Backpropagation, Gradient Descent, (Regression, SVM, Tree, KNN) models, K-Means Clustering.
• DL - DNN, CNN, Architectures (VGG, Residual, Inception, Efficient), Image Classification/Segmentation.
• Misc - Data Leakage, Cross-Validation, Augmentation, TensorFlow coding.
• Code & Slide - GitHub.

Journal/Conference Reviewer

• BMVC 2024 (British Machine Vision Conference)
• ISBI 2024 (IEEE International Symposium on Biomedical Imaging)
• IEEE Journal of Translational Engineering in Health and Medicine (IF - 3.4)
• Elsevier Biomedical Signal Processing and Control (IF - 5.1)
• Nature Scientific Reports (IF - 4.4)

Author / Contributor

• Keras/TensorFlow - Added Conversion to multibackend (JAX/PyTorch), DeBertaV3 classifier for MCQ task, GroupedQueryAttention, pad_images, and to_ordinal. Ongoing - MelSpectogram. (Code - GitHub)
• gcvit-tf Created from scratch in TensorFlow, explained in a live notebook, converted model weights from PyTorch, and created a live demo. (Code - GitHub | Notebook - Kaggle | Space - HuggingFace)
• TransUnet-tf - Created from scratch in TensorFlow. (Code - GitHub | Notebook - Kaggle)
• LLM-Science-Exam - Multi-backend model for scientific questions using Keras-NLP classifiers for GPU & TPU, with a live notebook for use-case. (Code - GitHub | Notebook - Kaggle)
• audio-cls-models - Conformer and ContextNet papers in TensorFlow, converted from STT to an Audio Recognition task, used for Fake Speech Detection with a live notebook. (Code - GitHub | Notebook - Kaggle)
• LLM-Detect-AI-Generated-Text - Demonstrated synthetic text detection with multi-backend `DebertaV3`. (Code - GitHub | Notebook - Kaggle)
• HuggingFace - Detected and resolved a discretized depth issue in the NYU-Depth-V2 dataset. (PR - GitHub)
• YOLOv5 - Identified and resolved bug related to best model checkpointing. (PR - GitHub)

Chairperson

• Organize IEEE EMBS Society's meetings, seminars, and workshops.
• Ensure member engagement and knowledge enhancement through society meetings.

Secretary

• Ensure member engagement and knowledge enhancement through society meetings.
• Assist speakers and guests during these events.

Class Representative (CR)

• Maintain class in absence of teacher, schedule quizzes, make announcements.
• Assist teacher in taking class seamlessly.
• Communicate as a channel between teacher and class.

Physics Olympiad

Math Olympiad

Math Olympiad

Science Fair