Faithful XAI
Multi-stage attention, SHAP, LIME, and Grad-CAM evaluation planning.
2027 MSc, MPhil and PhD Research Applicant
Trustworthy and Explainable AI for Real-World Decision Systems
I develop and evaluate interpretable machine-learning systems for agriculture, healthcare, and time-series applications. My work includes three peer-reviewed publications and a custom attention-based rice-leaf classifier evaluated on field images.
3.84
B.Sc. CGPA (Dean's x2)
3
Peer-reviewed publications
6.5
IELTS English Band
2027
Target research intake
Specimen 001
M. S. Sarker
Research Snapshot
Where the work sits.
Three directions, each backed by a specific project or publication — not a list of interests.
Robust agricultural vision
Mixed background rice-leaf dataset, group-aware splitting, custom lightweight multi-attention model.
Interpretable health and forecasting
Parkinson's thesis and the OLS/ARIMA climate paper.
Featured Research
Three Projects.
Project 01 — Agricultural Vision
Rice Leaf Disease Classification
A lightweight, edge-compatible multi-attention classifier for leaf-pathogen identification under mixed background conditions.
3 Mix
Datasets (Field & White)
5
Classes
1.47M
Params (Edge Ready)
82.33%
Test accuracy
0.7730
MCC
0.9576
ROC-AUC
Custom lightweight multi-attention based model optimized for resource-constrained edge computing devices, deploying channel and spatial attention mechanisms (320×320 input) trained on a fixed group-aware 70/20/10 split.
The training set blends three distinct datasets containing both natural field background photos and laboratory-controlled white background photos to maximize model generalization and robustness against environmental variation.
Experimental
Explanation figures and transfer-learning baselines in progress.
Code — private, available on request
Project 02 — Interpretable Medical AI
Parkinson's Disease Prediction
Detecting Parkinson's from speech features without sacrificing explainability.
756
Initial Features
6
Eval Datasets
10
Selected Features
98.28%
Ensemble Accuracy
Weighted
Ensemble Model
SHAP/LIME
XAI Pipelines
Problem: Speech-based detection models that are accurate but opaque about why they flag a patient.
Data: 188 patients, large high-dimensional dataset (756 columns). Applied 3 feature-selection methods and 2 balancing techniques to generate 6 distinct dataset variants, evaluating which combination best handles this large feature space.
Method: Reduced the 756 features down to the 10 most critical predictors, trained a weighted ensemble compared against seven individual models, and applied SHAP and LIME to interpret predictions.
Result: 98.28% accuracy from the weighted ensemble, gaining the highest performance across all models.
Under Review
B.Sc. thesis, manuscript under review.
See publication entry
Project 03 — Time-Series Forecasting
Climate Forecasting Baselines
OLS and ARIMA compared across 61 years of meteorological data, four countries.
61 Yrs
Historical Data
4
Countries
OLS/ARIMA
Tested Baselines
0.30
Bangladesh RMSE
Data: 61 years of historical temperature data across Bangladesh, Saudi Arabia, Japan, and Russia.
Method: hybrid ARIMA structures against OLS linear trends, with stationarity and residual testing.
Result: OLS reached RMSE 0.30 for Bangladesh temperature forecasting — a reminder that complex models need honest baselines before any novelty claim.
Published
IEEE COMPAS 2025.
IEEE Xplore
Publication Footprint
Research Evidence.
Peer-reviewed journal and conference work, with project and manuscript status stated separately.
Peer-Reviewed & Published
Published
Springer Journal Review
Artificial Intelligence Review (2026)
Leveraging explainable AI for sustainable agriculture: a comprehensive review of recent advances
Aditya Rajbongshi, Fatema Tuz Johora, Arafat Hossain, Md. Salauddin Sarker, Md Habibur Rahman, Md Wahidur Rahman, Fahad T. Alotaibi, Mohammad Ali Moni.
My Roledeep learning, machine learning, and XAI analysis synthesised into the review.
Abstract: This systematic review details current advances in Explainable AI (XAI) within precision agriculture. It synthesizes over 150 papers to map out local post-hoc attribution methods (LIME, SHAP, and Grad-CAM) across leaf pathogen detection, soil nutrient forecasting, and yield predictions. We propose a structured 4-tier taxonomy that evaluates model transparency, and discuss key challenges concerning explainability faithful bounds.
@article{rajbongshi2026leveraging,
title = {Leveraging explainable AI for sustainable agriculture: a comprehensive review of recent advances},
author = {Rajbongshi, Aditya and Johora, Fatema Tuz and Hossain, Arafat and Sarker, Md. Salauddin and Rahman, Md Habibur and Rahman, Md Wahidur and Alotaibi, Fahad T. and Moni, Mohammad Ali},
journal = {Artificial Intelligence Review},
volume = {59},
pages = {105},
year = {2026},
doi = {10.1007/s10462-025-11459-5}
}
Published
IEEE Conference Paper
IEEE COMPAS 2025
Bridging Climates: Weather Forecasting Using OLS and ARIMA Models with a Multi-Country Dataset
Arafat Hossain, Md. Salauddin Sarker, Aditya Rajbongshi, Most. Rakiba Khanom Jisa, Maria Afrin Bindu, Kayes Mohammad Abdullah.
My Rolemethodology, code implementation, and the results section.
Abstract: This research presents a comparative diagnostic study of standard statistical models—Ordinary Least Squares (OLS) regression and Autoregressive Integrated Moving Average (ARIMA)—for predicting historical climate indicators. We utilize 61 years of longitudinal meteorological records across Bangladesh, Saudi Arabia, Japan, and Russia. Standard residual tests confirm OLS yields lower prediction errors (RMSE 0.30) for baseline variables in agricultural forecasting.
@inproceedings{hossain2025bridgingclimates,
title = {Bridging Climates: Weather Forecasting Using OLS and ARIMA Models with a Multi-Country Dataset},
author = {Hossain, Arafat and Sarker, Md. Salauddin and Rajbongshi, Aditya and Jisa, Most. Rakiba Khanom and Bindu, Maria Afrin and Abdullah, Kayes Mohammad},
booktitle = {2025 IEEE 2nd International Conference on Computing, Applications and Systems (COMPAS)},
year = {2025},
publisher = {IEEE},
doi = {10.1109/COMPAS67506.2025.11381792}
}
Published
IEEE Conference Paper · 94.74% Accuracy
IEEE ICCIT 2025
Bridging Educational Gaps: Predicting Retakes Among Bangladeshi Undergraduates with Machine Learning and Explainable AI
Nusrat Zahan Nila, Md. Ali Mahmud Pritom, Fatema Tuz Johora, Aditya Rajbongshi, Md. Salauddin Sarker, Md. Ashrafuzzaman.
My Roleexplainable AI (XAI) analysis and the results section.
Abstract: Course failures and retakes severely interrupt academic timelines in higher education. We present a predictive model using records from 474 Bangladeshi undergraduates, deploying an ensemble soft voting classifier that achieves 94.74% risk-prediction accuracy. We incorporate SHAP and LIME explanatory pipelines to isolate core socio-economic and preparatory vectors responsible for academic risk.
@inproceedings{nila2025bridgingeducational,
title = {Bridging Educational Gaps: Predicting Retakes Among Bangladeshi Undergraduates with Machine Learning and Explainable AI},
author = {Nila, Nusrat Zahan and Pritom, Md. Ali Mahmud and Johora, Fatema Tuz and Rajbongshi, Aditya and Sarker, Md. Salauddin and Ashrafuzzaman, Md.},
booktitle = {2025 International Conference on Computer and Information Technology (ICCIT)},
year = {2025},
publisher = {IEEE},
doi = {10.1109/ICCIT68739.2025.11491491}
}
Under Review & In Progress
Under Review
B.Sc. Thesis · Medical AI · First author
B.Sc. Thesis Manuscript
Enhancing Parkinson's Disease Prediction Using Feature Selection Models (RF, RFE, and ANOVA) and Explainable AI: A Comparative Analysis of Machine Learning Performance
Md. Salauddin Sarker, et al.
My Roledataset collection, methodology design and implementation, and the results section — primarily code implementation.
Abstract: Speech disorder is a significant challenge for Parkinson's patients, impairing effective communication; early detection is essential to minimise movement issues and disability. While prior work detects Parkinson's with machine learning, a gap remains in the explicability of those models. We propose a system that both detects Parkinson's disease and analyses the reasons behind predictions to identify the actual causes, increasing transparency and trustworthiness. Using a medical dataset of 188 patients with a high-dimensional feature space (756 columns), we apply three feature-selection methods and two class-balancing algorithms (SMOTE and IHT) to generate six unique datasets to evaluate performance, subsequently reducing the feature set down to the 10 most critical predictors. An ensemble learning approach is compared against seven individual models; the weighted ensemble achieves the highest accuracy at 98.28%. Two explainable AI (XAI) techniques ensure transparency and interpretability of the chosen model. The proposed system is expected to play a vital role in the medical sector.
In Progress
Edge AI · Industrial inference · First author
Concept in Development
EdgeVision: Quantization-Aware Inference for Industrial Edge
Md. Salauddin Sarker, et al.
FocusProposed benchmarking study for resource-constrained industrial inference.
Concept: This planned work will evaluate post-training and quantization-aware methods for industrial vision models on resource-constrained hardware. Latency, model size, energy use, and accuracy retention will be reported after reproducible hardware benchmarking.
Applied Engineering Proof
Professional Experience.
Industry data-systems work that demonstrates the engineering rigor and reproducibility behind my research.
Senior Principal Officer — Software Developer (ERP)
CurrentLeading enterprise ERP and software development: designing Odoo modules, relational data models, and reporting workflows that turn raw operational data into auditable, analysis-ready systems. Also holds additional responsibility for mobile application development supporting the ERP platform.
Odoo
PostgreSQL
Python
ERP Architecture
Mobile Development
Executive — Software Developer (ERP)
Q Cosmetics Ltd · Dhaka, BangladeshBuilt the connective data layer between operations and decision-making: Odoo 17 ERP workflows, PostgreSQL ETL pipelines, and BI dashboards — converting raw transactional data into auditable, analysis-ready evidence.
35%
Faster order-cycle workflow via ERP rollout.
12×
Reporting cut from ~3 hours to ~15 minutes.
6+
Systems deployed across sales, inventory, purchase.
Odoo 17
PostgreSQL
ETL Pipelines
Python
BI Dashboards
The Pathway
Academic & research journey.
2019
Started B.Sc. Engineering
2022
First Dean's Award
2024
Thesis completed · 2nd Dean's Award
2025
First peer-reviewed publications
2026
Professional ERP role, Redmin
2027
Target research intake
Distinction & Engagement
Leadership & Recognition.
National STEAM Olympiad 2023
9th Nationally · Campus Ambassador
Selected campus leader for the national innovation drive; ranked top 10 nationally on technical pitch.
Vice President
UFTB STEAM Club
Led technical mentorship for 200+ members and organised engineering hackathons and innovation drives.
Executive Member
UFTB Robotics Club
Coordinated large-scale robotics competitions and cross-discipline engineering teams.
Japanese Secretary
UFTB Language Club
Drove multilingual academic exchange and Japanese (N5) learning initiatives.
Resources
Application portfolio documents.
Available
Available
Transcript (B.Sc.)
Research Statement
Research Directions
Open Questions.
What I want to investigate next — not claims that these are already solved.
Q1
How can explanation faithfulness be evaluated under dataset shift?
Field images vary in lighting, background, and acquisition device — a live question for the rice-leaf classifier's planned Grad-CAM evaluation.
Q2
How can agricultural vision models avoid background and acquisition shortcuts?
The rice-leaf dataset's field backgrounds make this concrete: a model can look accurate while keying on background texture rather than the leaf.
Q3
How can interpretable models remain reliable on small or imbalanced health datasets?
The Parkinson's project (188 patients, SMOTE/IHT balancing) is the starting point — the open question is how far that reliability generalizes.
The Foundation
A strong candidate for graduate excellence.
01
Academic Foundation
B.Sc. in Educational Technology & Engineering, CGPA 3.84/4.00, two Dean's Awards (2022 & 2024).
02
Research Evidence
Three peer-reviewed publications spanning explainable AI, deep learning, and time-series forecasting.
03
Research Readiness
Experiment design, model evaluation, XAI fidelity testing, and reproducible technical documentation.
04
Funding Objective
Seeking full tuition waiver plus research or teaching stipend for the 2027 academic year.
Technical skills
Python
PyTorch
scikit-learn
SHAP / LIME / Grad-CAM
SQL / PostgreSQL
Faculty Endorsements
Academic References.
Contact details are shared directly with universities and prospective supervisors upon request.
Farhana Islam
Assistant Professor & Chairman
Dept. of Educational Technology & Engineering, University of Frontier Technology, Bangladesh
Available on request
Aditya Rajbongshi
Assistant Professor
Dept. of Educational Technology & Engineering, University of Frontier Technology, Bangladesh
Available on request
Contact
Get in touch.
I am seeking fully funded MSc, MPhil, Master by Research, or PhD supervision for 2027. My primary interests are trustworthy XAI, agricultural computer vision, and medical AI.