Projects

Project 1

SixfinityOCR - Undergraduate Thesis

Developed an end-to-end Bangla OCR system by creating a custom annotated dataset, evaluating traditional OCR engines and Vision Language Models, and building a user-friendly application.

OCR NLP VLMs Python
Live Demo Github Ripo
Project 2

Innovest

Innovest is an integrated platform that streamlines fundraising, investing, events, community engagement, and real-time communication for startups and investors.

React Node.js Python Django
Live Demo Github Ripo
Project 3

Job Hunter

The Job Hunter System is a user-friendly web platform that connects job seekers with diverse opportunities while allowing companies to efficiently recruit qualified talent.

Html CSS PHP MySQL
Live Demo Github Ripo
Project 4

LFRSPC

LFRSPC Robot is an Arduino and ESP32-based autonomous line-following robot that cleans solar panels using a two-step water spraying and wiper mechanism, achieving 90–97% efficiency with minimal human intervention.

Arduino ESP32 IR Sensors
Demo Video Github Ripo

Research Projects

A Comprehensive Survey of Deep Learning-Based OCR and Information Extraction for Low-Resource Languages

Summary:
This review paper presents a comprehensive overview of the development of Optical Character Recognition (OCR), Document Layout Analysis (DLA), Information Extraction (IE), and Key Information Extraction (KIE), with particular emphasis on Bangla and other low-resource languages. It reviews the evolution of OCR from traditional approaches to modern deep learning, transformer, vision-language, and large language model (LLM)-based methods, while examining their datasets, methodologies, evaluation metrics, and applications.
The survey also analyzes research trends and identifies major challenges, including limited annotated datasets, complex Bangla scripts, handwriting variability, document degradation, inconsistent evaluation, and limited cross-domain generalization. It further discusses practical applications in legal records, finance, healthcare, historical archives, and government services. Finally, the paper highlights emerging directions such as self-supervised learning, few-shot learning, synthetic data generation, multimodal models, and LLM-based OCR refinement, providing a clear view of the current research gaps and future opportunities for building more accurate, robust, and scalable document understanding systems for low-resource languages.
Research Area:
Optical Character Recognition (OCR), Document Layout Analysis (DLA), Large Language Model (LLM) and Intelligent Document Understanding.
Status:
Manuscript in Preparation

A Carbon-Aware Software-Level Energy Performance Modeling Framework for AI-Driven Workloads with a Weighted Penalty Metric

Summary:
The paper proposes a Carbon-Aware Software-Level Energy Performance Modeling Framework for evaluating the sustainability of traditional and AI-driven workloads. It measures execution time, CPU utilization, memory usage, and power consumption across different workloads and input sizes, then estimates energy consumption and carbon emissions.
The collected data is preprocessed and used to train regression models, including Linear Regression, Random Forest, and MLP Regressor, for predicting sustainability-related outcomes. Finally, the proposed Weighted Penalty Metric (WPM) combines normalized carbon, execution time, CPU, memory, and delay factors to produce a unified sustainability score, where lower values indicate more sustainable execution.
Research Area:
Green Computing, AI Sustainability, Machine Learning-Based Energy and Carbon Modeling and Sustainable Software Evaluation Metrics.
Status:
Manuscript in Preparation

An End-to-End Unpaired CycleGAN Framework with the CUID Dataset for Underwater Image Enhancement

Summary:
This research proposes an end-to-end unpaired CNN–GAN framework for underwater image enhancement using the USIS dataset. The framework addresses common underwater degradations such as color distortion, low contrast, blur, and poor visibility using a CycleGAN-based architecture with adversarial, cycle-consistency, and identity losses.
The proposed framework is trained using unpaired underwater and enhanced image domains, eliminating the need for difficult-to-obtain paired ground-truth images. The USIS dataset combines diverse underwater scenes from different environments and sources to improve the model's ability to handle varying lighting, water conditions, and levels of degradation. The experimental results are evaluated through qualitative analysis and no-reference metrics such as UCIQE and UIQM, demonstrating improvements in color balance, contrast, sharpness, and overall visibility. In addition, systematic validation and ablation studies are used to examine the effectiveness of the framework's components. The resulting system provides a practical and reproducible foundation for future underwater vision applications, including object detection, marine monitoring, environmental analysis, and underwater robotics.
Research Area:
Digital Image Processing, Computer Vision, Image Restoration, Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs).
Status:
Academic Research Project

Adversarial Machine Learning Attacks on Intrusion Detection Systems: Analysis and an Improved Defense Pipeline

Summary:
The paper investigates the vulnerability of machine-learning-based Intrusion Detection Systems (IDS) to adversarial attacks using the NSL-KDD dataset. Five models SVM, Random Forest, Gradient Boosting, Extra Trees, and MLP are evaluated, with the MLP selected for further adversarial analysis. The study focuses on two gradient-based attacks, FGSM and PGD, to assess how small perturbations can affect IDS performance.
The experimental results show that the MLP achieves 77.16% accuracy on clean data, but its performance decreases considerably under adversarial conditions. At ε = 0.05, accuracy falls to 59.42% with FGSM and 65.78% with PGD, demonstrating that adversarial perturbations can significantly weaken IDS detection capability. The study concludes that high clean-data accuracy does not guarantee security in adversarial environments and emphasizes the importance of adversarial training, ensemble defenses, feature denoising, and robustness-aware architectures. Future work also suggests exploring broader robustness metrics and hybrid defense strategies.
Research Area:
Computer Security, Adversarial Machine Learning and Intrusion Detection Systems (IDS).
Status:
Academic Research Project

ShunoBondhu: A Voice Assistive App for Paddle Rickshaw Pullers in Bangladesh

Summary:
ShunoBondhu is a voice-assisted mobile application designed to address the safety, navigation, and financial challenges faced by paddle rickshaw pullers in Bangladesh. The app uses Bangla voice interaction to provide landmark-based navigation, voice-based daily earnings tracking, and emergency assistance through SOS voice commands or one-tap interaction. It follows a User-Centered Design (UCD) approach, incorporating formative research, prototyping, peer evaluation, and feedback from rickshaw pullers.
The study identifies that many rickshaw pullers have limited literacy and digital skills, making text-heavy applications and conventional map interfaces difficult to use. Through focus groups, ethnographic observation, interviews, prototyping, and real-world evaluation, the researchers found a strong preference for voice communication and landmark-based directions, along with the need for simple earnings tracking and quick emergency support. The prototype was refined with large icons, simplified navigation, clearer Bangla voice guidance, emergency contacts, and route-cost suggestions. The proposed future version will incorporate real-time GPS navigation, offline Bangla voice recognition, context-aware safety monitoring, automated SOS alerts, detailed income and expense analysis, and larger-scale field testing.
Research Area:
Human-Computer Interaction (HCI), Mobile Application Design, Accessible and Inclusive Technology and Urban Mobility.
Status:
Academic Research Project

Understanding Academic Resilience in Students with Depression : An Analysis Using Classification Algorithms

Summary:
This study investigates the use of machine learning techniques to predict academic resilience among university students experiencing depression. Using demographic, academic, lifestyle, and mental health factors collected from 600 students in Dhaka, Bangladesh, the research compares several classification algorithms, including Logistic Regression, Decision Tree, Random Forest, KNN, Naive Bayes, SVM, and ensemble methods.
The findings show that machine learning can effectively identify patterns associated with academic resilience, with Random Forest achieving the highest individual model accuracy of 82.22% and an ROC-AUC of 91.93%. Gradient Boosting and AdaBoost also demonstrated strong performance, highlighting the effectiveness of ensemble-based approaches. The study suggests that predictive models can support the early identification of students who may require academic and mental health interventions, helping educational institutions develop more personalized support strategies.
Research Area:
Machine Learning, Data Mining, Academic Resilience Prediction and Mental Health Analytics.
Status:
Academic Research Project