Imran Hossen
Master of Science by Research in Computer Science
University College Dublin, Ireland
Education & Certification
Master of Science by Research in Computer Science
Master's thesis
Bachelor of Science in Computer Science
Undergraduate Thesis
- Interpretability Method Development (State of the Art): Developed a SOTA interpretability method for vision transformers with 16.4% higher IoU and a 20% improvement in both precision and F1-score compared to existing methods. A blind test with 500 images and 5 heatmaps per image (4 existing vs. ours), 3 human evaluators preferred the proposed method in 84% of cases. Published a paper titled: F2HF: Feed Forward Network as noisy head filter in vision transformer interpretability.
- Research Supervisor: Prof. Kazi Saeed Alam
International English Language Testing System (IELTS)
Experience
Software QA Engineer
- Automation Framework Development: Developed an automation framework using playwright for internal use by the Software Engineers of the company saving up to 80% manual QA time.
- MVP Launch: Launched 3 MVP within 4 months, leading to more successful contract signed as offshore team.
Machine Learning Engineer
- Computer Vision Pipeline: Developed a machine learning (vision) pipeline for real-time tennis serve analysis by serve extraction, serve-type classification, and 3D racket trajectory reconstruction from 60 fps videos. Achieved ≤1° racket tilt accuracy and 20 cm ball bounce localization.
Research Assistant
- Multi-modal Aneurysm segmentation: Made a 3D ViT from scratch to detect Intracranial Aneurysm from multimodal (MRI, CT & MRA) data. Attained 86.3% AUC ROC score, comparable to SOTA. (Model open-sourced).
- Skin-lesion Segmentation: Developed & fine-tuned vision transformers for skin lesion segmentation (swin-transformer, vit enc etc.) to compare gradient shattering effect on generated mask. Outperformed SOTA by 3%.
Machine Learning Team Lead
- Reinforcement Learning in Robotics: Lead the design and implementation of a Reinforcement Learning based path-finding solution for the Mars environment, optimizing for: Remaining charge, Battery capacity, Solar light availability, Distance. Achieved top-10 rank globally in International Rover Challenge (IRC).
Projects
Virtual Dress Try-On Model (with GUI)
- Developed a Double-UNet–based diffusion model for virtual try-on, achieving an FID score of 9.01 (almost SOTA). Separate UNets for the cloth and person–cloth generation architecture aligns cloth on body more accurately.
Autism Screening and Therapy Platform
- Developed a 3D Transformer for behavioral video analysis, achieving 95.2% accuracy on just 300 data samples. Model deployed on AWS Lambda as a serverless function for automatic scaling. Backend deployed on EC2.
Skills
Programming Languages: Python, SQL, C++, Typescript.
Machine Learning Libraries: Pytorch, Transformers, SciPy, Scikit-learn.
Programming Frameworks: MLFlow, OpenCV, Pillow, Spring, Django, Vue.JS, Langchain.
Cloud and Databases: Docker, AWS S3, AWS EC2, AWS Lambda, MongoDb, MySQL, PostgreSQL, Redis.