Authors - Akshar Sodankoor, Avanish Shenoy, B Monish Moger, Mohnish Gowda, Prafullata Kiran Auradkar, Subramaniam Kalambur Abstract - Virtualization is essential for efficient resource utilization in cloud and development environments. With the growing adoption of gRPC as a Remote Procedure Call (RPC) framework, evaluating its performance across different virtualization technologies has become crucial. This work benchmarks the performance of four gRPC call types: unary, client-streaming, server-streaming and bi-streaming, across four lightweight virtualization technologies. Docker, gVisor, Firecracker and nanos unikernel. The analysis examines CPU utilization, memory utilization and network capabilities to provide a comprehensive comparison. The results show that docker delivers the best performance across all metrics. Firecracker shows comparable latency performance to docker, but consumes higher memory. Nanos unikernel exhibits CPU utilization similar to that of docker, but has the highest latencies in all cases except unary gRPC call. gVisor exhibits the lowest CPU utilization under heavier workloads and also has the lowest latencies for client-streaming and server-streaming gRPC calls.
Wednesday August 26, 2026 9:30am - 11:30am IST Virtual Room DGOA, India
Authors - Aryan Goyat, Aditya Maan, Vimmi Malhotra Abstract - Deep learning has transformed artificial intelligence and enabled major breakthroughs in applications like computer vision, natural language processing, medicine, cybersecurity, and robotics. Through the use of deep neural networks, it enables automatic feature learning and exceeds machine learning-based methods in accuracy and flexibility. Challenges including excessive computational expense, uninterpretable nature, and ethics are still major hurdles to its widespread application. This article discusses the development and applications of deep learning and presents new research directions that seek to overcome its limitations. Federated and decentralized learning methods improve security and privacy by enabling collaborative model training without raw data sharing. Explainable AI (XAI) techniques, including SHAP and LIME, enhance the interpretability of deep learning models, making their decision-making more transparent. In addition, energy-efficient deep learning methods, such as model pruning, quantization, and neural architecture search (NAS), are being designed to minimize computational and environmental expenses. The emergence of self supervised learning further minimizes dependence on labeled data, making deep learning more feasible across domains. Future developments will center on the fusion of deep learning with reinforcement learning, symbolic AI, and evolutionary algorithms to build more generalizable and efficient systems. These technologies will power the next wave of intelligent, ethical, and sustainable AI solutions.
Wednesday August 26, 2026 9:30am - 11:30am IST Virtual Room DGOA, India
Authors - Anita Agrawal, Ruhi Panjwani, Aditya Mallik Abstract - This paper explores the design and implementation of a multi-user, multi-access web application tailored specifically for automated weather stations (AWS). By examining real-world scenarios and user interactions, we identify key design considerations, including system performance, security, and scalability. The study aims to provide practical insights for developers to create efficient and user-friendly web applications that effectively handle large-scale weather data and cater to various user access levels.
Authors - Krishna Shirsath, Abdullah Ansari, Riyaz Memon, Phiroj Shaikh Abstract - Real-time collaboration is essential for modern software development, enabling developers to work together seamlessly from different locations. This paper presents DevTogether, a collaborative coding platform that facilitates efficient teamwork through live code editing, customizable collaboration sessions, integrated chat, a collaborative drawing board for design prototyping, and real-time video meetings powered by WebRTC. The platform incorporates an intelligent code assistant using the Gemini API, along with an autosave mechanism to ensure workflow continuity. Built using the MERN stack, DevTogether emphasizes scalability, low latency, and responsive performance while addressing synchronization conflicts and communication challenges. Experimental evaluations and simulated performance metrics underscore its effectiveness compared to similar platforms.
Authors - Arnav Shukla, Subhashree Choudhury, Logeshwaran R. Abstract - One of the key limitations of any crowdfunding platform that uses blockchain technology is the lack of transparency regarding fund usage by campaign creators after receiving donations. The current paper addresses this issue. For this investigation, we conducted a detailed examination of the relevant literature on decentralized applications (DApps) and the use of smart contracts to automate administrative tasks. We identified a significant gap in existing platforms: the ambiguity surrounding post-donation and how fund management can be unfair, which blockchain alone does not address effectively. Our findings highlight the strengths of blockchain security and automation capabilities, which ensure a safe and trustworthy environment for users. Our proposed solution integrates a decentralized voting model and a collusion prevention algorithm called EigenTrust to empower donors with a participatory role in decision-making processes, eliminating the need for centralized authority. In this study, we implement and evaluate a decentralized voting model that uses specific techniques to prevent any attacks or chances of misuse of powers that would then empower donors to participate in critical campaign decisions, enhancing trust and satisfaction by allowing them to verify the responsible use of their contributions. By reducing uncertainty around fund allocation, our model increases a more engaging and end-to-end secured donor experience, encouraging donations and supporting the long-term success of social crowdfunding projects. In all, this paper presents a novel approach to increase crowdfunding platforms using blockchain technology with a voting model paired with collusion prevention to address the issue.
Authors - P S Sree Harsha, Harshitha S, Ayush Sisodiya, M P Deepti, Sarasvathi V Abstract - Concerning the Conventional recruitment methods, which has many challenges such as the poor matching of the candidates to the right requirements, issues of transparency and issues of privacy in dealing with sensitive information. This leads to the hiring of candidates who are not qualified to meet their requirements and compromise the organizational data. To solve these problems, Human Resource Blockchain Intelligence Recommendation System (HRBIRS) is proposed to use Blockchain and Recommendation system technologies for getting an efficient recruitment process. Blockchain is evolving technology providing transparent and secure data. Privacy issues are resolved by decentralized architecture. The Hybrid recommendation system makes recruitment effective by determining the qualifications, skills and experience of the job seekers relevant to certain job recruitments posted by the HR. One of the most attractive features of HRBIRS is the peer approval and endorsement systems within the organization without bias to a particular candidate. This paper provides detailed information on the architectural design, implementation and its performance features and demonstrates how this system could be beneficial for recruitment processes throughout the organization by providing data security and enhancing the effectiveness of selecting the right candidate.
Authors - Arunangshu Giri, Dipanwita Chakrabarty, Manash Routray Abstract - The present study emphasized on how enrollment of loyalty programs at fuel retail outlets get enhanced through digital communication and intelligent promotional strategies. The study has evaluated the three major dimensions like participation intention and promotional efficiency to understand the efficacy of the loyalty programs organized by different fuel retail outlets. 454 Indian customers were interviewed over a three months period using a structured questionnaire. Cross-sectional descriptive research design was followed for the study. Both qualitative and quantitative analysis was done using NVivo and SPSS-28 software. The study revealed that customer participation in loyalty programs was highly influenced by flexible redemption options, digital reward system and exclusive benefits. Again, the study has shown how staff knowledge, promotional policies, customer loyalty and digital engagement influence promotional effectiveness. The study acknowledged the pivotal role of intelligent communication strategies in enrichment of customer adaptability towards digitized loyalty programs. Establishing a seamless communication between the customers and digital platforms along with effective staff training can be prudential for optimal customer engagement, sustainability and loyalty.
Authors - Lalithya Govardhan, Shalini M S, Gagan Deep P S Abstract - This work shows an extensive multimodal system of mood detection and customized playlist recommendation based on EEG, GSR, and face recognition. Brainwave activity for emotional evaluation is sensed by EEG electrodes, while GSR sensors provide skin conductance, heart rate variability, and temperature values for physiological behavior. Facial behavior with emotional facial expressions is determined through facial landmarks. Preprocessing entails Butterworth filters for EEG frequency bands, GSR data normalization, and facial feature extraction from a pretrained model (FER) to track eyebrow position, mouth curvature, and eye openness. EEG features are examined using frequency domain analysis, whereas GSR and facial features are classified using Random Forest. To increase precision, a fusion model aggregates predictions by weighted averaging or majority voting, with EEG assigned the greatest weight due to its high correlation with mood. After determining the emotional state, a suitable playlist is suggested: energetic songs for happiness, relaxing music for stress, comforting songs for sadness, and relaxing music for relaxation. This feature-based recommendation system enhances personalization through the use of features like tempo, genre, and mood to provide a dynamic and interactive listening experience for the user
Authors - A. Revathi, A. Sunidhar Reddy, Geetika Alapati, R. Pranay Abstract - This paper presents the performance of the grocery identification system concerning Telugu grocery items, considering both native and non-native speakers. Speech recognition for Telugu groceries presents a unique challenge due to variations in pronunciation, accent, and noise conditions. This study explores the implementation of a Gaussian mixture model (GMM) classifier in conjunction with rasta-perceptual linear prediction (RASTA-PLP) features to enhance the accuracy of Telugu grocery identification. Rasta-PLP effectively captures robust speech features by suppressing unwanted spectral variations, while GMM provides a probabilistic framework for classification. The proposed system is trained on a dataset comprising commonly used Telugu grocery names and evaluated under diverse acoustic environments. Experimental results demonstrate improved recognition performance, showcasing the effectiveness of RASTA-PLP in feature extraction and GMM in classification. This work contributes to developing efficient speech-based interfaces for regional language applications, facilitating voice-driven grocery identification systems. The recognition accuracy of the proposed system is approximately 99%, ensuring high reliability in real-world applications. This technology benefits society by aiding visually impaired individuals and non-Telugu speakers in grocery identification, enhancing accessibility and convenience. By enabling seamless voice-based interaction, promotes inclusivity and improves social equity through technological advancement.
Authors - A.Revathi, Reethikaa Vallinayagam, S. Sivaranjani, A. Deepthi Abstract - This research work introduces a system for identifying genuine speech and recorded (replay) speech through Mel-Frequency Cepstral Coefficients extraction and uses k-means clustering for classification purposes. Speech features obtained from various speakers undergo normalization procedures before receiving cluster assignment during training sessions. During the testing phase speaker identification depends on measuring the distance between input features against cluster centroids. The confusion matrix indicates system performance by showing correct genuine speech detection through high diagonal values yet exhibiting lower off-diagonal values to indicate possible attacks based on recorded speech. Auto-correlation together with cross-correlation serve to evaluate the similarities between speakers. Strong recognition of the same speaker is indicated by high auto-correlation values but weak cross-correlation values demonstrate effective differentiation between different speakers. The AVSpoof dataset serves as the experimental foundation because it includes ten recording subjects who are distributed between five male and five female speakers. The acceptance and accuracy evaluation for the system happens through testing samples which proves its ability to recognize genuine speech from recordings as well as identify distinct speakers properly.
Authors - Moushmee Milind Kuri, Ganesh Pathak Abstract - Cyberbullying is a growing concern across social media platforms, necessitating advanced detection mechanisms to mitigate its impact. Traditional machine learning models often struggle with understanding contextual dependencies and ensuring model interpretability. This paper proposes a hybrid deep learning approach that combines BERT and RoBERTA for feature extraction and Bidirectional Long Short-Term Memory (BiLSTM) networks for sequential dependency learning. To enhance interpretability, attention mechanisms such as Self-Attention and Bahdanau Attention are integrated, allowing the model to focus on crucial words contributing to classification. The proposed system aims to improve accuracy, scalability, and explainability while addressing key challenges in cyberbullying detection. This research lays the groundwork for developing more transparent and effective AI-driven moderation systems for online safety.
Authors - Devi V S, Durgalashmi C V Abstract - This study assesses the effectiveness of income tax deductions and exemptions in promoting savings and investments in India. The Indian government has implemented various tax incentives to encourage individuals to save and invest, including provisions under sections 80C, 80D, and others. These deductions and exemptions are designed to stimulate economic growth by fostering long-term financial planning among individuals. The research examines the impact of these provisions on individual taxpayers' behavior and their overall influence on savings and investment patterns. Through a comprehensive analysis of available data, the study identifies the key tax incentives that have led to increased savings in instruments such as Provident Funds, National Savings Certificates, and insurance products. Additionally, the research evaluates the extent to which these tax benefits contribute to fostering a culture of investment and financial security. The study concludes that while tax deductions and exemptions have provided some incentives for savings, their effectiveness is often limited by lack of awareness and financial literacy. To further promote savings and investments, the study recommends improvements in policy communication, accessibility, and the alignment of tax incentives with broader economic goals.
Authors - Akash Tibeli, Saroja V Siddamal, Suneeta V Budihal Abstract - The AHB to APB Bridge is crucial component in System-on-Chip (SoC) designs, Achieving efficient communication between the pipelined AHB bus and the non-pipelined APB bus. In the proposed work a AHB to APB bridge is built using a bridge architecture which enables to translate pipelined, burst-oriented, high speed AHB transactions into sequential, low-power APB transactions by maintaining synchronization and data integrity. It was developed with a FSM to manage transactions and pipelining to maintain efficiency. Verification was performed using a Universal Verification Methodology testbench environment through direct and random testcases of burst, single, sequential, non-sequential transactions. 80 testcases were tested to obtain a functional coverage of 88%.
Wednesday August 26, 2026 12:30pm - 2:30pm IST Virtual Room DGOA, India
Authors - Judy K George, Elizabeth Sherly Abstract - Convolutional Neural Networks are extensively employed in critical domains such as computer vision, medical imaging, and autonomous systems. Enhancing model interpretability by providing users with concise and context-relevant explanations of CNN decision making such as visualizing feature maps or saliency regions, enables a deeper understanding of the model’s internal representations and inference process. The proposed work presents a deep learning framework integrating a ResNet-based U-Net architecture with a Fixation Point Generator (FPG) to perform classification and saliency aware reconstruction on the hand-written dataset. The model leverages transfer learning by employing a pre-trained ResNet-18 as the encoder backbone, enabling robust feature extraction. A custom decoder reconstructs input images while a classification head predicts digit labels. To enhance model interpretability, a Fixation Point Generator predicts spatial attention maps (saliency maps) from high-level global features, highlighting regions of interest that influence model decisions. This implementation aims to bridge the gap between classification performance and model explainability, offering insights into the model’s focus areas through learned attention. The model got an accuracy of 98.44 on the Malayalam handwritten dataset, 97.81 on English handwritten dataset, and 99.56 on the MNIST dataset.
Authors - Meghali Kalyankar, Om Pratap Gajra, Prathamesh Vilas Sagvekar, Mehul lalit Sharma, Zoheir Shahid Shaikh Abstract - Deepfakes are an emerging threat to digital authenticity and security, hence a proper detection technique needs to be created in order to establish public study confidence. A thorough roadmap to the development of deepfake detection software has been provided in this paper, reviewing the state-of-the-art algorithms, such as XceptionNet, EfficientNet, and hybrid models integrating spatial and temporal analysis. It provides methodologies for implementation, data preprocessing, and software pipeline development, serving as a practical guide to researchers and developers. Theoretical study to application-oriented practice closes the gap in terms of bottom line development and adaptive detection systems addressing the growing menace of deepfake media.
Authors - PRANAY SAMAL, K R LOKESH KUMAR, CHAKILELA SAIRAJ, G VIDYA SRI, SUSHAMA RANI DUTTA, SUKLA SATAPATHY Abstract - Music plays a significant role in our day-to-day life, and selecting appropriate songs can enhance the experience. This paper describes an intelligent music recommendation system that applies machine learning to recognize what users prefer and recommend music that they will like. It incorporates various approaches, including considering user decisions and music attributes, to enhance suggestions. The system adapts based on user actions and refines recommendations over time. Findings indicate that this method provides easier and more precise music discovery. This paper emphasizes how technology can assist in providing a higher quality and better personalized music experience.
Authors - Arya Tripathi, Akash Mecwan Abstract - In recent years, the RISC-V architecture has emerged as a promising platform for embedded systems, offering flexibility and open-source accessibility. Consequently, the demand for secure communication in embedded devices, particularly within the Internet of Things (IoT) ecosystem, has driven the adoption of cryptographic algorithms. Integrating cryptographic functionalities into RISC-V architecture presents unique challenges, requiring innovative solutions to optimize performance and security. In response, the proposed design introduces an approach to address these challenges by incorporating a dedicated cryptoprocessor module into the RISC- V architecture specifically designed to handle encryption and decryption tasks efficiently. The cryptoprocessor module employs the Blowfish-64 algorithm to ensure robust security while lowering the computational overhead. Blowfish is a well-established symmetric-key block cipher known for its simplicity and efficiency. The compact design of the cryptoprocessor module significantly reduces resource utilization and execution time compared to existing implementations while preserving security and functionality. The design emphasizes low resource utilization, achieving a utilization rate of 34% (11,340 out of 33,216 available units) with an execution time of 160 ns. The implementation is carried out using Verilog HDL for the Cyclone II EP2C35F672C6 based FPGA.
Wednesday August 26, 2026 12:30pm - 2:30pm IST Virtual Room DGOA, India
Authors - Kirti Karande, Sujata Kadu, Deven Shah Abstract - Breadth-First Search (BFS) is a foundational graph traversal algorithm, it’s systematic layer-by-layer exploration of nodes makes it invaluable for a variety of domains, including transportation networks, social network analysis, and artificial intelligence. However, traditional BFS implementations face challenges when dealing with large-scale graphs due to memory limitations and inefficiencies in handling massive datasets. This project addresses these challenges by integrating BFS with a CSV-based data storage system, enabling efficient traversal of large graphs without relying on a traditional SQL database or requiring the entire graph to be loaded into memory. The graph data, comprising nodes and edges, is stored in CSV files, which act as lightweight and accessible storage. The implementation is memory-efficient due to the use of Pandas DataFrames for handling CSV data and NetworkX graphs for traversal. Additionally, the integration of a machine learning model from Scikit-learn, a memory-efficient library, ensures effective prioritization of edges without excessive computational overhead. In this project, we address a key limitation of the traditional Breadth-First Search (BFS) algorithm: its inability to consider edge weights during traversal. It is unsuitable for scenarios where varying edge weights significantly impact the traversal outcome, such as in shortest-path calculations for weighted graphs. To overcome this drawback, our project integrates a machine learning (ML) model to analyze and prioritize edges based on their weights, effectively augmenting BFS for weighted graphs. By leveraging CSV-based storage and combining it with an ML-driven edge prioritization mechanism. this project offers a scalable solution for managing and analyzing large, weighted graphs. This combination ensures that the navigation system not only computes the shortest path but also suggests the most practical and efficient routes tailored to user preferences or constraints.
Authors - Dipak Ligade, Chiranjit Das, Rupali Parte, Masira Kulkarni, Shivraj Jadhav, Abhishek Mohite Abstract - —This project aims to automate tree counting and forest land diversion assessment through satellite image combined with advanced computing techniques. The treatment of forest re sources needs accurate monitoring because growing environmental challenges such as deforestation, biodiversity loss, and climate change require it for sustainable land management. The research uses satellite imagery along with machine learning and deep learning tools, specifically convolutional neural networks (CNNs), to precisely detect and count trees across expansive territories. The study demonstrates how satellite analytics technologies will enhance forestry applications with their capabilities for better tree enumeration at higher efficiency and greater accuracy.
Authors - Suchanta Ravan, Prashant Dhotre Abstract - Conventional identification techniques that depend on privacy concerns and credentials are becoming more vulnerable to web-based risks like hacking and data breaches. The need for sophisticated authentication techniques has grown dramatically because of identity theft, cyberthreats, and illegal access. Traditional security methods, such as PINs and passwords, are insufficient for high-security applications since they are vulnerable to phishing, brute-force assaults, and credential breaches. To improve safety and tackle problems like privacy threats, spoofing, and accessibility problems, this study suggests a strong adaptive authentication mechanism that integrates biometric along with behavioral assessment. The multimodal authentication framework guarantees a smooth and easy verification process while also enhancing security. This structure guarantees a smooth and safe authenticating process by utilizing cutting-edge security methods like encryption, machine learning, and multifaceted biometrics in conjunction with a user-centric architecture. By combining behavioral biometrics with conventional authentication techniques, total authentication reliability is increased, and cyber risk is mitigated. The effectiveness of the suggested approach in lowering susceptibility to cyberattacks while preserving superior usability and consumer satisfaction is demonstrated by experimental findings, Highlighting the importance of two-way authentication.
Authors - S M Boomika, C M Tulasi, Sharvani V Nagur, Bhagyashri Badakali, Nalini C Iyer, Preeti Pillai, Ujwala Patil Abstract - LiDAR and cameras play a vital role in autonomous vehicles by providing complementary data for object detection and environmental perception. However, achieving seamless data integration from these sensors depends on partial and temporal synchronization. Unlike conventional methods that depend on pre-calibrated datasets, our methodology utilizes a custom-acquired multimodal dataset comprising both image and video data from a monocular camera and point cloud data from a VLP-16 Velodyne LiDAR sensor. In this paper, we proposed a comprehensive framework for LiDAR and camera calibration and temporal synchronization of real time data, synthesized and validated in a controlled lab environment. Calibration of the raw data was performed using a checkerboard as the target to ensure accurate spatial alignment between heterogeneous sensor systems.The collected corpus is further timestamped, synchronized, and validated.The accuracy of the proposed methodology is evaluated by projecting LiDAR points onto image frames, enabling qualitative verification of spatial and temporal consistency. The proposed method integrates target-based calibration with software-level timestamp synchronization to create a reproducible and scalable calibration pipeline. Results demonstrate accurate alignment across modalities, validating the effectiveness of our approach. This 1 work provides a practical contribution to multi-sensor fusion research, especially for applications requiring custom datasets or operating in constrained environments.
Authors - Wendrila Biswas, Arunangshu Giri, Dipanwita Chakrabarty, Dibyendu Rath Abstract - The study has examined the effect of user engagement (UE), perceived benefit (PB), and perceived risk (PR) of wearable sensor-based healthcare devices adoption. User empowerment (UEM) in IOT-enabled healthcare has been explored on the basis of two established theories, Technology Acceptance Model (TAM) and Behavioral Reasoning Theory (BRT). A cross-sectional online survey was conducted from November 2024 to January 2025 involving 361 valid Indian respondents and the collected responses were analyzed through NVivo software for qualitative analysis. SEM (structural equation modeling) was done for quantitative analysis and hypothesis testing. The findings have shown a positive association between UE and PB and between UE and PR. Again, the study has revealed that PB and PR positively influenced UEM. The study contributes both to existing literatures and making managerial decisions by establishing how benefits from wearable sensor-based healthcare devices can be explored by avoiding the perceived risk of the consumers and how they can get empowered with the same.
Authors - Vaishali Langote, Siddhesh Kulkarni, Aaditya Ghorpade, Aditya Songirkar, Aditya Chincholkar Abstract - Identifying customer retention is essential for decreasing lost revenues as well as maintaining an established base of loyal customers. By reviewing historical data that includes customer demographics, purchasing habits and behaviours, businesses will be able to determine which customers are going to discontinue using their services or products. In generating models that can identify customers at risk, this process includes machine learning models such as decision trees, logistic regression and neural networks. It is important that predictive retention can work provided the right algorithms are selected, and reliable data is sourced. Continual updates and improved models will enhance accuracy, giving firms the opportunity to keep up with changes in how consumers behave. The models will also give businesses the ability to produce more targeted retention marketing plans since they will not only identify at-risk customers but also give clear data on what they are doing to create customer churn.
Authors - Sharon Koshy, Padmadas Sundaram Abstract - The intensifying depletion of natural resources, fueled by world population growth and unsustainable consumption, poses severe threats to global sustainability. Specifically, the ICT and smart infrastructure industries make substantial contributions to resource inefficiencies through growing e-waste, inefficient material recovery, and unsustainable construction methods. Forecasts suggest that by 2050, with a projected 9.8 billion world population, resource use will surpass planetary limits, urging rapid interventions in resource management and the transition to circular economies. Despite growing recognition, inefficiencies in recycling infrastructure, defective waste-to-energy technologies, and inadequate water management persist to drive global resource insecurity and further environmental degradation. Solutions must be backed by evidence-based policy design, technological development, and systemic change. In this context, the combination of Artificial Intelligence and biomimicry offers a new way to increase sustainability and resilience in systems. AI-based models improve resource efficiency, reduce environmental footprint, optimize waste management, facilitate predictive maintenance, and enhance material recovery, while biomimicry offers nature-inspired solutions for sustainable design, energy efficiency, and waste reduction. These technologies not only foster resource recovery but also set the stage for the creation of wiser, more sustainable industries and cities. In conclusion, this study high- lights the revolutionary power of ICT that AI and biomimicry make possible to create closed-loop, self-sustaining models that boost urban resilience, sustainability, and efficiency, maximize recovery of resources, minimize waste, and maximize value for a truly circular future.
Authors - Priya Surana, Sushma Vispute, Madhura Kalbhor, Shubhangi Vairagar, Pragati Ugale, Imtiyaz Syeda, Mahek Yakumsha, Ashish Suryawanshi Abstract - This research presents a YouTube Comments Analyzer that leverages machine learning and deep learning algorithms to examine and classify user comments. A large volume of comments is processed by the system, enabling it to detect key patterns, including sentiment classification and emotion detection. Using natural language processing and machine learning techniques, the tool provides meaningful insights to content creators for understanding their audience and to moderators for identifying problematic content. Researchers can also benefit by studying online commentary at scale. Our team collected video comments from various genres to train and develop the models, followed by evaluation using multiple performance metrics. The analysis tool achieves 96% accuracy in sentiment detection and 90% accuracy in emotion detection, successfully identifying complex patterns that manual evaluation often misses. To demonstrate the practical applicability of our models, we further developed a web-based application that integrates the analysis pipeline, providing an accessible platform for real-time comment analysis. This research highlights the effectiveness of automated text analysis in social media environments and demonstrates real-world applications for YouTube content management and audience engagement strategies.
Authors - Omkar Kalantre, Jyoti Joglekar Abstract - Optimized Bounding box fitting around an object is necessary for accurate localization of the Region of Interest (ROI), so that features extracted from the ROI are useful for many computer vision applications. Current methods tend to be inefficient, imprecise, and with high computational complexity. In this work a novel algorithm is presented that is designed for fitting a bounding box around an object that covers maximum part of the object as ROI,. The improvement in inserting bounding box enhances the process of recognizing, tracking, and classifying objects, which is highly valuable for applications such as surveillance, autonomous driving, and security. In this work we are proposing a novel algorithm for fitting a bounding box around an object to maximize the object area covering and for minimizing the background clutter as a part of ROI.
Authors - Mohan Sellappa Gounder, Rohan Mahantesh Kamatgi, Sharath Prabhu T M, Sanya Gupta, Seema Abstract - This research investigates the application of the DINO (Distillation with No Labels) framework, a self-supervised learning approach, for efficient road and pothole segmentation. By integrating a DINO-enhanced ResNet-50 backbone with a U-Net model, this study addresses segmentation challenges in dynamic environments. The framework employs momentum encoders, multi-crop training, and stability mechanisms to facilitate robust feature extraction without requiring labeled datasets. Through strategic fine-tuning, the model achieves precise segmentation of road surfaces and potholes, making it a promising approach for real-world applications in autonomous systems and infrastructure assessment. This study further discusses model evaluation, comparison with state-of-the-art approaches, and its implications for transportation infrastructure.
Authors - Gopal D. Upadhye, Ranjana Jadhav, Aryan Pungale, Ashish Shadija, Nikita Rajput, Pranav Pendse Abstract - A data-informed system is described for generating crop recommendations and crop yield forecast based on a variety of data sources of farmer-level soil characteristics, historical crop yield records, and meteorological variable data. In the proposed system, crop recommendations based on a classification algorithm and crop yield estimates based on a regression algorithm are provided to farmers. The data-driven crop recommendations and crop yield forecasts will improve decision-making by providing the farmer with data-based recommendations providing the productivity isolation. The data-informed system will utilize machine learning algorithms to process the data and analyze the complex interaction of the various farming agri-parameters in the farm operation. Composition of soil nutrient values, weather patterns, and historical productivity variable data will be a key ingredient in the model to provide farmers with singularly specific crop selections. Ability to yield prediction gives farmers anticipate yield of the crops, improve resource planning. The validation tests demonstrate better accuracy than traditional heuristics, improving farmer overall risk reliability and increasing efficiency, sustainability. The results shows us that the transformative role of machine learning in agriculture and the associated movement toward precision farming practices
Authors - Vanshika R Kavi, Sujata Kotabagi Abstract - Semiconductor production demands high-quality control to detect faulty wafers early on in the production process. Manual inspection and rule-based systems are conventional methods that are time consuming and error-prone. This research investigates machine learning (ML) based wafer detection on a dataset of 590 sensor readings per wafer, with wafers being labeled as good (+1) or faulty (-1). Several traditional ML models, such as Logistic Regression (LR), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Random Forest, are tested for defect classification effectiveness. The processing of data includes handling missing values by dropping features with high missing data and using median imputation. Feature selection is done through SHAP (Shapely Additive Explanations) analysis and correlation filtering to select only the most important sensor readings. Feature scaling is done to maintain consistency in data distribution. For handling the class imbalance in the dataset, SMOTE (Synthetic Minority Over-sampling Technique) is employed to create synthetic samples for the minority class to enhance model learning. Once trained, the models are evaluated on the basis of accuracy, precision, recall, F1-score, confusion matrix, and SHAP-based explainability analysis. SVM and Random Forest perform better compared to other models with 97-99% accuracy, and KNN does not perform well because of high dimensionality. The research showcases how ML is able to automate defect detection, increase production efficiency, and minimize human inspection errors. Work for the future encompasses ensemble learning optimization, real-time deployment, and semi-supervised learning optimization for enhanced defect classification in the semiconductor industry.
Authors - Piyusha S. Shetgar, Asha V. Thalange, Rohini R. Mergu, Aishwarya Khobare Abstract - Throughout the world, the number of educational institutions has significantly increased in recent decades. But the majority of recently established universities continue to manage their resources, including their hostels, using traditional methods. These conventional methods are frequently hindered by innate restrictions that negatively impact the organization's overall effectiveness. This study suggests an automated hostel lodging management system that is made with Microsoft Access as the underlying database and Visual Basic as the programming language to handle these issues. To stop unwanted access, the system has an integrated authentication algorithm. The system that has been built leverages face recognition technology to address the shortcomings of conventional approaches. It provides a graphical user interface, dependability, efficiency, and improved security by implementing access control mechanisms.
Wednesday August 26, 2026 3:30pm - 5:30pm IST Virtual Room DGOA, India
