I’m Rafi Majid, an AI researcher and software engineer working on efficient deep learning, multimodal AI, model compression, medical intelligence, and code intelligence.
My work focuses on reducing computational cost while preserving useful representations — and on understanding when model predictions should actually be trusted.
SELECTED WORK
Current and selected projects across model compression, medical vision-language learning, code intelligence and low-resource NLP.
EFFICIENT AI · KNOWLEDGE DISTILLATION · ACTIVE RESEARCH
Confidence-Gated Structural-Contrastive Knowledge Distillation for compressing transformer models while preserving useful teacher representations.
MEDICAL AI · MULTIMODAL · RELIABILITY
A decoupled reliability monitor for chest X-ray/report alignment under hard cross-patient mismatch and representation drift.
CODE INTELLIGENCE · SECURITY · MODEL COMPRESSION
CodeBERT-based source-code security analysis with knowledge-distillation strategies for lightweight vulnerability detection.
BANGLA NLP · TOKENIZATION · LANGUAGE TOOLS
A Bangla + Banglish tokenizer with language-aware normalization, grapheme handling and SentencePiece Unigram tokenization.
PUBLISHED RESEARCH
The study compares recurrent deep-learning architectures for predicting cyclone trajectories using historical Bay of Bengal data, with emphasis on practical forecasting performance.
Majid, R., Hasan, A., Mostarin, S., Alam, K.R., & Rahman, R.M. (2024). “A Deep Learning-Based Study for Cyclone Track Forecasting: Comparative Analysis Using Historical Data from the Bay of Bengal.” Springer Nature.
DOI: 10.1007/978-3-031-53274-0_1FORECASTING WORKFLOW
Conceptual illustration of the forecasting setup — not a reproduction of experimental results.
PUBLICATIONS
Three research outputs spanning low-resource vision-language learning, cyclone forecasting and graph/hypergraph analysis.
PREPRINT · GRAPH & HYPERGRAPH LEARNING
A leakage-controlled, matched-k comparison of graph and penalized hypergraph spectral clustering.
ARXIV · MULTIMODAL · LOW-RESOURCE AI
LLM-assisted translation refinement for Bengali vision-language research.
SPRINGER · SCIENTIFIC ML · TIME SERIES
Comparative analysis using historical Bay of Bengal cyclone data.
EXPERIENCE
A technical path combining academic research with practical software systems.
Developing backend and full-stack systems with Spring Boot, React and MySQL, alongside Unity/C# simulation environments.
Supported undergraduate teaching while building research experience across machine learning, AI and software systems.
Developing CoGS-KD, TRUST-VL-CXR and transformer-based code intelligence work focused on efficiency, representations and reliability.
EDUCATION & TECHNICAL FOUNDATION
GRADUATED 2023
North South University · Dhaka, Bangladesh
CGPA 3.49 / 4.00Research experience developed across deep learning, multimodal learning, scientific forecasting, graph learning and transformer-based code intelligence.
RESEARCH INTERESTS
Tools: Python, Java, C++, PyTorch, TensorFlow, Hugging Face, CUDA, Git, Linux and LaTeX.
ALGORITHMIC PRACTICE
GET IN TOUCH
I’m open to research collaboration, PhD opportunities, and conversations around machine learning, multimodal systems, model compression and code intelligence.