Md Sajeebul Islam Sk

I'm an Artificial Intelligence researcher at BRAC University in Dhaka. I did my master's in Computer Science and Engineering from BRAC University, where I was supervised by Dr. Md. Golam Rabiul Alam and bachelor in Mathematics from Khulna University, where I was supervised by Dr. Md. Rafiqul Islam.

Email  /  Google Scholar  /  LinkedIn  /  GitHub

Research

I'm interested in artificial intelligence, machine learning, computer vision, deep learning, generative AI, NLP, world model, and similar fields.

Shows why stochastic LLM rankings can be uncertain under unknown coupling.

SafeTube-WM is an occupancy-input world model that connects future occupancy prediction with trajectory-level collision risk and uncertainty estimation.

Audits cultural bias in LLM benchmarks and proposes culturally responsive evaluation.

RegFire-WM is a one-step wildfire world model that separates next-day prediction, transition diagnostics, and error scoring and examine where it succeeds, fails, and supports planning.

An evidence-aware English–Bengali financial Question Answering (QA) framework based on Qwen3-8B with 4-bit QLoRA, lightweight evidence selection, Bengali-digit normalization, number-preserving Bengali and code-mixed augmentation, and concise calculation supervision.

FedDP-PALD is a privacy-preserving framework that enables hospitals to jointly synthesize medical data from chest X-ray and ECG signals, employing differential privacy and latent diffusion to safeguard patient records. It resists membership-inference attacks while preserving diagnostic accuracy.

CauSAIL is an AI system that predicts maritime supply chain disruptions by combining ship tracking, satellite images, and text, as well as using specific mathematical techniques for LLM bias mitigation.

Reducing the difficulty of accurate and explainable Lumbar Spinal Stenosis diagnosis from MRI by proposing an end-to-end explainable VLM framework using Spatial Patch Cross-Attention, Adaptive PID-Tversky loss, and automated radiology report generation module.

Addressing sparse, irregular, and zero-inflated long-horizon livelihood forecasting by developing a mask and time-gap-conditioned Transformer state-transition world model.

Propose a hybrid model that reduces noise, biased prediction, overfitting, poor generalization, and model size by using Morlet–DB4 wavelet denoising, ADASYN class balancing, pretraining CBAM–GAT–TST attention architecture and predicts Normal/Abnormal ECGs and five cardiac diseases from ECG diverse datasets.

Create Bengali language based speech datasets and a speech-text based personality classification framework by introducing MoMF feature extraction, and developing DistilRo (DistilBERT and RoBERTa) and BiG (Bi-LSTM and GRU) soft-voting ensemble models to predicts ten different personality traits.

Analyze 5-second speech signals with Haar, Sym2, db4 wavelets and calculated PSNR, and SNR values for both drug-addicted and non-addicted people.

Experience

Research Associate
Advisor: Dr. Tanvir Hasan

Research Assistant
Advisor: Dr. Md. Ashraful Alam

Research Assistant
Advisor: Md. Mehedi Hasan Shawon, Dr. Md. Golam Rabiul Alam

Research Assistant
Advisor: Dr. Md. Golam Rabiul Alam

Education

Thesis: Unveiling Personality Traits through Bangla Speech using Morlet Wavelet Transformation and Soft-Voting Classifier

Thesis: Identification of Drug-addicted and Non-addicted People Using Voice Signal Through Haar Wavelet Transform

Miscellanea

Template stolen from Yijia Weng (and also Jon Barron).
Last updated: May 10, 2026