Academic & Applied Research

Research & Publications

Investigating few-shot learning, vision-language architectures, and reliable AI behaviors under domain shift. Committed to developing training-free frameworks and rigorous empirical safety benchmarks.

Academic Focus & Trajectory

Few-Shot Learning, Vision-Language Models & AI Safety

I am a Computer Science & Engineering graduate (CGPA 3.92/4.00, Magna Cum Laude) with over four years of production software engineering experience and an active research track record. My primary research investigates how to build reliable, sample-efficient AI models under distribution shifts and how parameter-efficient adaptation (PEFT/LoRA) impacts safety boundaries.

Current Focus: Seeking research-based Master's opportunities to advance work on multimodal reliability, failure-mode taxonomy, and clinical abstention.

Research Metrics at a Glance

Conference Paper (DICTA 2026)Accepted
Systematic Review (Medical VLMs)In Progress
Independent Safety Study (LoRA)In Progress
Undergraduate Thesis (NLP)86.7% IMDB

Showing 5 of 5 items

Accepted · DICTA 2026 (September 2, 2026)Lead Author

Conference Paper · 2026

Training-Free Whitened Local-Global Fusion for Few-Shot Plant Disease Recognition

2026 International Conference on Digital Image Computing: Techniques and Applications (DICTA 2026), Australia

Supervised by Md. Ismail Hossen, Griffith University, Australia (Google Scholar, LinkedIn)

A fully training-free framework on a frozen ResNet-18 that combines ZCA-whitened local descriptors with a global Mahalanobis metric to bridge the laboratory-to-field domain gap on the PlantVillage benchmark.

Key Contributions & Highlights:

Eliminates gradient-based retraining, enabling immediate few-shot diagnosis on edge agricultural hardware.
ZCA whitening decorrelates local visual descriptors to suppress irrelevant background foliage noise.
Combines global Mahalanobis distance metric with local patch alignment to generalize across lab-to-field domain shifts on the PlantVillage dataset.

Computer Vision
Few-Shot Learning
PlantVillage
ResNet-18
ZCA Whitening
DICTA 2026

In ProgressCo-Author

Systematic Review · 2026

Few-Shot Fine-Grained Medical Diagnosis with Vision-Language Models: Taxonomy, Benchmarks, and Open Challenges

Systematic Review & Meta-Analysis in Few-Shot Medical Vision-Language Modeling

Supervised by Md. Ismail Hossen, Griffith University, Australia (Google Scholar, LinkedIn)

Conducting a comprehensive systematic review and failure-mode taxonomy of few-shot medical vision-language models, establishing standardized benchmarks for localized visual attention and clinical abstention in rare disease classification.

Key Contributions & Highlights:

Failure-mode taxonomy of medical vision-language architectures evaluated on scarce pathology datasets.
Benchmarking clinical abstention mechanisms when model uncertainty exceeds diagnostic thresholds.
Evaluating attention map localization to ensure clinical explainability and safe alignment.

Vision-Language Models
Medical AI
Few-Shot Learning
Systematic Review
Multimodal
AI Safety

In ProgressLead Author

Empirical Study · 2026

Large-Scale Safety Evaluation of Public LoRA Modules

Independent Research

Examining how LoRA-based parameter-efficient fine-tuning impacts safety and alignment in large language models. Evaluating 100+ public modules for toxicity, jailbreak susceptibility, and demographic bias amplification.

Key Contributions & Highlights:

Empirical safety benchmarking across 100+ public Hugging Face LoRA adapters.
Analyzing safety boundary shifts and representation drift introduced through low-rank adaptation.
Proposing auditing heuristics for enterprise LLM adapter deployments.

LLM Safety
LoRA
PEFT
Alignment
Evaluation

Completed · 2024Lead Author

Undergraduate Thesis · 2024

Advancing NLP Tasks Through Fine-Tuning of Foundation Language Models

Undergraduate thesis, supervised by Sharfuddin Mahmood, Assistant Professor, AIUB

Supervised by Sharfuddin Mahmood, American International University-Bangladesh (AIUB)

Fine-tuned GPT-2 Medium with adaptive few-shot learning and Optuna hyperparameter optimization. Achieved 86.7% accuracy on the IMDB sentiment benchmark under limited labeled data.

Key Contributions & Highlights:

Automated hyperparameter exploration using Optuna Bayesian optimization.
Demonstrated fast few-shot convergence on sample-scarce sentiment tasks.
Completed as undergraduate thesis majoring in Software Engineering (Magna Cum Laude).

NLP
GPT-2
Optuna
Few-Shot Learning
Transformers

Completed · 2022Author

Academic Project · 2022

Machine Learning for Clinical Risk Prediction

Academic Research Project, AIUB

Supervised learning study on early stroke risk prediction using imbalanced clinical data. Benchmarked SVM, Random Forest, and Naive Bayes; identified Random Forest as most robust under imbalance (86.5% accuracy, 10-fold cross-validation 0.87).

Key Contributions & Highlights:

Handled severe clinical class imbalance with specialized resampling and metric tuning.
Evaluated feature importance curves on critical physiological biomarkers.

Machine Learning
Healthcare AI
Classification
Imbalanced Data

Advisors & Research Collaborators

Grateful to collaborate with and be mentored by respected researchers and academic faculty.

Md. Ismail Hossen

Supervisor & Co-Author

PhD Candidate & Teaching Fellow at Griffith University, Australia. Former Lecturer in Computer Science, AIUB.

Google Scholar·LinkedIn

Sharfuddin Mahmood

Thesis Supervisor

Assistant Professor & Special Assistant [OSA], Department of Computer Science, American International University-Bangladesh (AIUB).

smahmood@aiub.edu

Dr. Md Shakhawat Hossain

Academic Referee

Associate Professor & PI, AI for Medical Imaging (AIM) Lab, Kochi University of Technology, Japan. Former Assistant Professor, AIUB.

md.shakhawat@kochi-tech.ac.jp

Interested in Research Collaboration?

I am eager to contribute to research projects on few-shot learning, vision-language models, and AI safety benchmarks. Let's connect and discuss ideas.

Reach Out for Collaboration Email Direct