Bengal College of Engineering & Technology

Traditional labor management systems face significant challenges, including delayed wage payments, worker exploitation, and difficulty in verifying employment history. These challenges are particularly common among migrant and freelance workers, who tend to suffer from non-payment and insecurity of employment. Moreover, the current systems are not transparent, and it is hard to monitor project progress, payments, and contractual terms, resulting in common conflicts between employers and workers. This paper introduces a Decentralized Labor Management System (DLMS) that utilizes blockchain technology. DLMS overcomes these by employing smart contracts, where employers first deposit funds before the commencement of work, with automatic payment at the end. This removes delays in wages and risks of non-payment, most especially for migrant and freelance laborers. It also securely stores worker profiles, job histories, and credentials on an immutable blockchain ledger, which allows for verifiable proof of experience and ability. The platform increases transparency by enabling real-time monitoring of projects, payments, and agreements, minimizing fraud and disputes between employers and workers. In times of crisis, like the COVID-19 pandemic, it can enable real-time location and condition monitoring of workers to help organizations and governments offer timely support. Through trust building, removal of financial barriers, and simplification of employment procedures, DLMS enables a more equitable and efficient labor system that favors both workers and employers.

Diabetic Retinopathy (DR) is a major global cause of preventable blindness, making early detection essential for timely treatment. Automated grading remains challenging due to severe class imbalance, high similarity between adjacent severity stages, and limited expert-labelled datasets. This work proposes CKD-Net, a three-stage progressive chained knowledge distillation framework in which knowledge is transferred from one model to the next $(\mathrm{M} 1 \rightarrow \mathrm{M} 2 \rightarrow \mathrm{M} 3)$. The framework uses an attention-enhanced ResNet18 backbone, CLAHE preprocessing, WeightedRandomSampler for class imbalance, and a cosine annealing learning-rate scheduler for stable convergence. Evaluated on a multi-class DR dataset of 2,750 fundus images across five severity stages, CKD-Net achieved 83% test accuracy and a Quadratic Weighted Kappa of 0.9078, indicating strong agreement with expert grading. The model also achieved 95%precision for healthy cases and 93% recall for proliferative DR. Index Terms-Diabetic Retinopathy, Knowledge Distillation, Attention Mechanism, ResNet, CLAHE, Class Imbalance, Medical Image Classification.

Pitaya or dragon fruit is a tropical fruit prized for its unusual look and several health advantages including a high vitamin and antioxidant content. Dragon fruit is becoming more and more popular worldwide which benefits emerging countries like Bangladesh, Vietnam, China, Thailand, Indonesia, Israel and India economically. The present study offers a very large collection of high resolution dragon fruit photos that will help the machine learning models in their determination of the ripeness and quality of the fruit. The dataset was created with the utmost care and guidance from specialists over a period of four months from three different locations in Bangladesh. The collection is of great importance in facilitating the operations of dragon fruit production by giving resources for robotic harvesting, quality evaluation, and packing systems etc. We performed quality assessment using the Nasnet Mobile, DenseNet121 and MobileNetV2 models which yielded test accuracies of 94.89

The growing expansion of the Internet of Things and wireless sensor networks has created an urgent demand for compact and reliable radio frequency energy-harvesting circuits. This study introduces the design, simulation and extensive performance of a high-efficiency single band radio frequency detection system optimized for 1.8 GHz operation. The detector is realized on a Rogers RO4003C substrate and employs the SMS7630-079LF Schottky diode, selected for its excellent detection capability and economic viability. The introduction of this filtering stage effectively suppresses undesired harmonic components produced during the rectification process, thereby improving the sensitivity and overall power conversion efficiency of the system. The circuit shows a sensitivity of 1.8 mV for every dBm through its simulation tests. The system shows increased sensitivity to 2.2 mV/dBm because of the band stop filter implementation. The system reaches its peak power conversion efficiency of 65.28% at a 1.5 kΩ load, which makes it suitable for applications that require low-power energy harvesting. These combined attributes establish the developed 1.8 GHz detector as a strong candidate for next-generation energy harvesting modules, self-powered sensor networks and intelligent embedded computing platforms within the expanding domain of the Internet of Things.

Sentiment analysis is pivotal for extracting insights from user-generated content across multilingual digital platforms. While traditional methods perform well in monolingual settings, they often struggle with the complexities of linguistic diversity, including syntactic variations, data scarcity in low-resource languages, and cross-lingual domain shifts. This study addresses these challenges by proposing a hybrid deep learning framework that synergizes the strengths of transformer-based models, sequential learning, and graph-based reasoning for robust multilingual sentiment analysis. Leveraging the Amazon Multilingual Review Dataset—spanning English, German, French, Spanish, Japanese, and Chinese—it evaluates standalone models (mBERT, BiLSTM, GNN) and novel hybrid architectures (BERT-BiLSTM, GNN-BERT) enhanced with attention mechanisms. In experiments demonstrate that hybrid models consistently outperform standalone approaches, with GNN-BERT achieving a 93.4% F1-score by effectively integrating contextual embeddings, sequential dependencies, and relational structures. Key contributions include: (1) a scalable framework for language-agnostic sentiment classification, (2) rigorous per-language evaluation revealing performance disparities (e.g., 92.2% F1 for English vs. 84.5% for Arabic), and (3) solutions for low-resource language challenges through cross-lingual transfer. The study also highlights the role of attention mechanisms in improving interpretability by identifying sentiment-relevant tokens. These advancements bridge critical gaps in multilingual NLP, offering practical applications in e-commerce and social media analytics. For future work, it outlines directions including zero-shot transfer learning, handling code-switching, and integrating multimodal sentiment analysis to further broaden real-world applicability.