Authors - Madan Kumar Sharma, Nadir Kamal Salih Idries, Abdullah Said Alkalbani, Satyanarayana Degala, Gopal Rathinam, Ankit Sharma Abstract - Multi-resonance (MR) Microwave sensors have emerged as a promising sensing device for mineral-based materials characterization due to their high sensitivity and precision. This research presents a novel microwave sensor for mineral-based material characterization. The sensor's resonating structure comprises a 3 × 3 array of circular-shaped complementary split-ring resonators (CSRR) coupled with four rectangular defected structures etched around the CSRR array. This unique configuration enhances electromagnetic interaction with the material under test (MUT), leading to precise characterization based on S-parameter analysis. To validate the sensor’s efficacy, simulation-based investigations were conducted on the mining-based materials, including chrome, copper, and quartz. The obtained results demonstrate distinct resonance shifts and attenuation variations corresponding to each mineral, highlighting the sensor’s capability to differentiate and analyze their dielectric properties. The proposed MR-sensor design provides a robust and efficient method for non-destructive material characterization, offering potential applications in the mining industry, quality control, and geophysical exploration.
Authors - Geethu Lakshmi G, P. Nagaraj, P. Chinnasamy Abstract - Lung cancer detection constitutes a paramount process in the diagnosis and management of one of the predominant contributors of cancer-induced humanity worldwide. The significance of early screening is underscored by its essential role in enhancing survival rates through the identification of disease during a stage amenable to treatment. Diagnostic methodologies, including imaging modalities, are routinely utilized for diagnosis. Furthermore, advancements in the realm of molecular biology have facilitated the emergence of biomarkers and genetic examines, thereby enabling a more accurate identification of lung cancer. The prompt and defined detection of lung cancer facilitates timely therapeutic interventions, which significantly influence both the efficacy of treatment and overall outcomes for patients. The primary objective of this research is to develop an optimization-based hybrid deep learning methodology for the detection of lung cancer. The initial phase involves pre-processing of input images through techniques such as color space transformation, data augmentation, resizing, and normalization. Subsequently, features derived from Slime Mould Algorithm-based Convolutional Neural Network (SMA-CNN) are employed for the detection of lung cancer, with CNN being trained utilizing SMA, extracted from pre-processed images. Finally, the Squeeze-Inception V3 model, which integrates SqueezeNet and Inception V3, leverages SMA to train the classifier. Consequently, the proposed SMA-based hybrid SqueezeNet-Inception V3 is utilized to classify instances as normal or abnormal. Empirical results designate that SMA-based hybrid SqueezeNet-Inception V3 attained an accuracy of 97.3%, a specificity of 96.1%, and a sensitivity of 98%, thereby underscoring its efficacy in the detection of lung cancer.
Authors - Narayan Gupta, Pawan Kumar, Prince Kumar Singh, Priyabart Kumar, Parampreet Kaur Abstract - For investors, accurately predicting stock market prices is a critical part of financial analysis. This work studies a wide variety of machine learning algorithms specifically designed for the task of predicting stock prices using a variety of techniques and the latest technologies available as of the time of study. The study critically compares a suite of algorithms, including Support Vector Regression (SVR), Random Forests, Decision Tree models, and Long Short-Term Memory (LSTM), each with differing strengths. Moreover, it investigates a variety of approaches that are focused on understanding the complex connections found in the past price data. The dataset used for this study includes extremely long-term stock price representatives over a time range from 2010 to 2024 for Tata Consultancy services (TCS), containing a wide range of information, including opening and closing trading prices, trading volumes, and a variety of calculated indicators reflecting market behaviour. To understand the first experiment carried out for this study, it can be seen that in combination mode, for the best-performing model, Random Forest has the best interpretability. In addition, both Support Vector Regression (SVR) and Decision Tree algorithms deliver both impressive short-term prediction results and clear explanations behind their decisions.
Authors - Sandeep Shinde, Samarveer Moray, Aditya Sakhare, Prathamesh Salokhe, Kedar Sathe Abstract - The exponential growth of digital documents, particularly PDFs, presents significant challenges in efficient information retrieval and extraction. Traditional methods often struggle with the complexity and variability inherent in large PDF documents. Recent advancements in Natural Language Processing (NLP), especially Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG), offer promising solutions. This paper presents a comprehensive system for efficient information extraction from large PDFs using RAG and LLMs. We propose a robust and scalable pipeline addressing challenges such as document segmentation, dynamic retrieval, and response contextualization. Through extensive experiments across multiple domains—including legal analysis, technical documentation, and scientific literature—we demonstrate that the proposed methodology significantly outperforms existing approaches in terms of accuracy, scalability, and efficiency. Our research lays the groundwork for integrating RAG and LLMs in various domains, offering a valuable tool for extracting knowledge from complex documents.
Authors - Rahul Pethe, Parag Puranik, Abhay Kasetwar Abstract - Wireless Sensor Networks (WSNs) have been designed and developed by numerous researchers over time, evolving based on emerging demands and environmental conditions. Over the years, various enhancements and improvements have been proposed, with multiple protocols introduced to address design challenges. In parallel, Mobile Ad Hoc Networks (MANETs) have witnessed tremendous growth in this wireless era. While many researchers have focused on protocols like DSR and hybrid approaches for energy-efficient clustering, these solutions have often proven to be time-bound and context-specific. To overcome these limitations and enhance the performance and lifetime of WSNs, we propose a new scheme based on the AODV (Ad hoc On-Demand Distance Vector) protocol within a mesh networking framework. Our approach achieves optimal results, including 100% throughput, minimal jitter, and significant improvements in network life-time.
Authors - Pranav Mittal, Prerna, Dhruv Bansal, Nikhil Panwar, Krish Tyagi Abstract - Stock market prediction is an intricate process in financial analysis, as its main aim is to predict the price trends and thus help to make trading strategies. But the stock market is so unpredictable with various factors affecting it, and thus predicting whether the value will rise or not becomes a difficult task. This study aims to predict the stock market prices with historical data that will be achieved via Deep Learning techniques in particular LSTM networks. Due to its strength in learning long-term dependencies and preserving the sequence information, LSTM (one of the variants of RNNs) is ideal for time-series data. Our approach leverages a dataset of daily stock prices from various financial indices over multiple years. The data is preprocessed using normalization techniques to improve model accuracy. The LSTM model is then compared to a traditional feed-forward neural network to demonstrate the superiority of LSTM in predicting shortterm stock trends. The models are optimized using the Adam optimizer, The results indicate that LSTM significantly outperforms the conventional models in forecasting accuracy. Moreover, this research introduced a hybrid LSTM-CNN method to extract features and prediction firmly. This study will contribute to financial forecasting by utilizing deep learning techniques and real trading scenarios. The research work was carried out using the programming language known as Python, deep learning tools such as TensorFlow and Keras, data management libraries such as Pandas and NumPy, and data representation software such as Matplotlib. It was found that the LSTM model has a much higher level of success when time series patterns are approximate than any other type of neural network, hence stock price predictions are more accurate. This study helps in how LSTM networks can be useful for forecasting in finance hence would be helpful to traders and market analysts.
Authors - Sheetal Phatangare, Komal Potdar, Yash Mahajan, Mandar Pandagale, Vivek Nikam Abstract - University students frequently encounter challenges in retrieving relevant academic information due to the limitations of traditional search engines. This research introduces KnowledgePilot, the first 1-bit Large Language Model (LLM) specifically designed to support university students. Leveraging the BitNet b1.58 architecture, which employs ternary parameterization (-1, 0, 1), KnowledgePilot achieves high performance with reduced computational costs, making it both resource-efficient and fast. The system integrates Retrieval Augmented Generation (RAG) pipelines, enabling it to access external academic data sources, thus minimizing hallucination issues common in LLMs and providing accurate, context-specific responses. The research also encompasses the development of tools for file conversion, dataset creation, model pretraining and fine tuning. Comprehensive evaluations will measure the system’s performance and user satisfaction, demonstrating its potential to significantly enhance student access to academic resources, while setting the stage for future advancements in low-bit AI technologies for education.
Authors - Yash Chavan, Arnav Sonawane, Arpit Pattiwar, Aditya Nagdive, Kaushalya Thopate Abstract - In today's fast-moving world, consumers rely on packaged foods. This is extremely important for easy access to detailed and personalized nutritional information. This project focuses on developing mobile applications for barcode scanning. This includes extensive food details, including ingredients, nutritional value, allergen warnings, and personalized consumption recommendations based on a person's health. Applications written with Python and Kivy provide a seamless user experience, allowing individuals to scan barcodes and upload images of ingredients to extract and analyze related information. Additionally, it includes optical character detection (OCR) using Tesseract, which extracts text from photos to allow users to analyze the ingredient list and nutritional name, even if barcode scans are not possible. By taking into account user nutritional limitations or illnesses such as diabetes, lactose intolerance, or gluten sensitivity, this application provides tailor-made health advice and helps individuals make found food decisions appropriately. A secure user authentication system improves the experience by storing your preferences and receiving recommendations created by tailors. The main goal of this project is to enable consumers to choose food in real time and promote healthier consumption habits. The combination of barcode scanning, OCR, and a structured database causes applications to close the gap between the complexity of food indicators and user understanding. Future improvements include mechanically learning-based ingredients, integration into real-time product databases, and expansion of several platforms beyond Android. This initiative represents an important step in using technology to improve consumer health awareness and ensure safer and sounder decisions for food consumption.
Thursday August 27, 2026 12:30pm - 2:30pm IST Virtual Room BGOA, India
Authors - Evangeline R C, Krupa Nirmal, Laasya P, Aishwarya K Abstract - Real-time pedestrian trajectory prediction is essential for enhancing safety and urban mobility, particularly in dense and dynamic environments. This paper introduces a video data processing system that accurately predicts pedestrian movement by analyzing sequences of video frames in real time. The system effectively handles challenges such as overlapping individuals, partial occlusions, and diverse walking behaviors, making it suitable for real-world deployment. The architecture is designed to be both modular and scalable, allowing for seamless integration into various applications such as traffic management, urban planning, and pedestrian safety enhancement. A user-friendly interface provides real-time visualization of the predicted trajectories, enabling accessibility for both technical and non-technical stakeholders, including urban planners and public safety officials. Extensive experiments conducted on multiple diverse datasets demonstrate the system’s reliability and accuracy across various conditions, including crowded scenes and irregular pedestrian movement. The system successfully captures complex behavior patterns and provides predictive insights that can help reduce pedestrian-related accidents. This research contributes significantly to the field of intelligent pedestrian monitoring systems. By combining real-time responsiveness with accurate trajectory prediction, the proposed system supports the development of smarter and safer urban infrastructure, fostering proactive decision-making and improved pedestrian safety in modern cities.
Authors - Nisha Dubey, Randeep Singh Abstract - Facial profile classification and acknowledgment have different applications in security, perception, and identity affirmation. This paper proposes a novel approach utilizing Convolutional Neural Frameworks (CNNs) to classify and recognize facial profiles. The proposed system utilizes a significant CNN designing to remove solid highlights from facial profiles, taken after by a classification layer to recognize profile classes (e.g., cleared out, right, frontal). The illustrate is ready on a tremendous dataset of facial profiles and finishes tall accuracy in classification (95.2%) and affirmation (92.5%) errands. Test comes around outline the system's quality to assortments in lighting, pose, and expression. Besides, the proposed system outflanks existing techniques in facial profile classification and affirmation. This work contributes to the movement of facial examination development, engaging its course of action in real-world applications such as identity affirmation, get to control, and perception. In particular, facial profiles provide an interesting challenge due to the distinctive variety of lighting, attitude, expression and disorders. Furthermore, large data records and accessibility of arithmetic violations have made it possible to prepare violent .CNN models for facial profile classification and detection
