Kazi Hafiz Md Asad
Hi, I’m Kazi Hafiz Md. Asad 👋
M.S. Researcher in Computer Science and Engineering
North South University, Bangladesh
Research Identity
I am an M.S. researcher working at the intersection of efficient NLP, large language model compression, trustworthy AI, and higher-order machine learning.
My research focuses on building machine learning systems that achieve strong performance while reducing computational cost, memory requirements, and reliability risks.
My current interests include:
- Efficient and trustworthy NLP
- LLM compression and parameter-efficient fine-tuning
- Knowledge distillation
- Quantization and memory-efficient training
- Graph and hypergraph representation learning
- AI for software security
- Multimodal healthcare AI
- Sustainable AI
Research Highlights
Ultra-Efficient LLM Fine-Tuning under Constrained VRAM
Qwen2.5-7B-Instruct
- Investigated memory-efficient fine-tuning using 4-bit quantization, QLoRA, adaptive low-rank optimization, and gradient-aware strategies.
- Conducted 27 controlled experiments analysing quality, GPU memory, runtime, and convergence stability.
- Developed layer-streamed QLoRA optimization achieving 63% peak GPU memory reduction while maintaining comparable performance.
TRUST-VL-CXR: Reliable Vision-Language Monitoring for Chest X-Ray Alignment
- Developed a decoupled reliability monitoring framework for medical vision-language models.
- Combined DenseNet-121 and BiomedBERT representations with an independent compatibility monitor.
- Achieved 0.7666 AUROC for hard image-report conflict detection under patient-disjoint evaluation.
CoGS-KD: Confidence-Gated Structural-Contrastive Knowledge Distillation
- Developing efficient code vulnerability detection models through knowledge distillation.
- Exploring CodeBERT-family teachers, structural learning, confidence gating, and contrastive objectives.
- Building reproducible experiments on PrimeVul for security-focused language models.
Graph vs Hypergraph Protein-Complex Recovery
- Studying higher-order biological representation learning.
- Developed leakage-controlled comparisons between graph and hypergraph spectral clustering.
- Evaluated biological coherence using statistical validation and enrichment analysis.
Education
North South University
M.S. in Computer Science and Engineering
2024 - Present
GPA: 3.55/4.00
Thesis: Ultra-Efficient Fine-Tuning of 7B-Parameter Large Language Models under Constrained VRAM using Dynamic Progressive Compression and Adaptive Stride Training
B.S. in Computer Science and Engineering
2019 - 2023
GPA: 3.53/4.00
Graduated Cum Laude
Thesis: BERT Knowledge Distillation for NLP Tasks
Publications
Larger Models Yield Better Results? Streamlined Severity Classification of ADHD-Related Concerns Using BERT-Based Knowledge Distillation
PLOS ONE, 20(2), e0315829, 2025
🔗 Paper · LastBERT Model
- Introduced LastBERT, a 29M-parameter student model distilled from BERT-base.
- Achieved 3.8× parameter reduction while maintaining competitive performance across GLUE benchmarks and ADHD concern severity classification.
- Demonstrated that carefully designed compact models can achieve strong performance without relying solely on larger architectures.
Strengthening False Information Propagation Detection: Leveraging SVM and Sophisticated Text Vectorization Techniques in Comparison to BERT
IEEE International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN), 2025
- Compared classical machine learning approaches with transformer-based models for misinformation detection.
- Evaluated SVM classifiers using TF-IDF, Word2Vec, and bag-of-words representations against BERT-based approaches.
- Analysed predictive performance versus computational cost for resource-efficient NLP deployment.
Preprints & Manuscripts
When Does Higher-Order Representation Improve Protein-Complex Recovery? A Leakage-Controlled, Matched-k Comparison of Graph and Penalized Hypergraph Spectral Clustering
Preprint, 2026
- Developed a leakage-controlled benchmark comparing graph and hypergraph representations for protein-complex recovery.
- Evaluated nine IntAct MITAB datasets using matched-k comparisons, statistical testing, null models, and Gene Ontology enrichment.
- Investigated when higher-order biological interactions provide measurable advantages over conventional graph projections.
Research Engineering
Selected implementations:
- QStride / AdaStride: Efficient LLM adaptation under constrained VRAM
- TRUST-VL-CXR: Medical vision-language reliability monitoring
- CoGS-KD: Efficient vulnerable-code detection
- LastBERT: Lightweight transformer compression
- Graph and hypergraph biological modelling
Technical Toolkit
Programming
Python, SQL, PHP, Linux, Git
Machine Learning
PyTorch, Hugging Face Transformers, scikit-learn, TensorFlow
NLP / LLM
BERT, RoBERTa, Qwen, GPT-family models, LoRA, QLoRA, RAG
Research Methods
Ablation studies, controlled benchmarking, statistical testing, leakage prevention, reproducible experimentation
Research Philosophy
Strong Baselines
+
Controlled Experiments
+
Transparent Evaluation
+
Reproducible Systems
=
Reliable AI Research
Connect
GitHub: https://github.com/donnowhattodo
Email: kazi.asad@northsouth.edu
LinkedIn: kaziasadcse
Google Scholar: Kazi Hafiz Md. Asad
Last updated: August 2026