Menglong Guo 1, Yu Sun 2 , 1 The Chinese University of Hong Kong, Ma Liu Shui, Hong Kong, 2 California State Polytechnic University, Pomona, CA 91768
This study presents a systematic comparison of LLM-powered database interfaces versus traditional SQL systemsfor inventory management, implementing two parallel Flask backends—a SQLite-based system using SQLAlchemyORM and an LLM-based system using DeepSeek to process natural language commands against JSON storage—with identical REST API endpoints enabling controlled comparison [10]. Experimental results reveal significanttrade-offs: the SQL backend achieved 12ms mean latency and 100% operational accuracy, while the LLM backendaveraged 1,850ms latency (154x slower) with 88% accuracy that degraded to 72% for complex multi-stepoperations. These findings demonstrate that while LLM-powered databases offer unprecedented query flexibility andnatural language accessibility, they currently incur substantial performance and reliability penalties; traditionalSQL systems remain superior for mission-critical applications requiring deterministic behavior and ACIDcompliance, while LLM approaches suit scenarios prioritizing user accessibility and dynamic query capabilitiesover guaranteed correctness and response speed.
Large Language Models, SQL Databases, Natural Language Interfaces, Inventory Management.
Ulugbek Shernazarov 1, Rostislav Svitsov 1 , 1 Bin Shi , 1 Telecom SudParis, France
Fine-tuning large language models for domain-specific tasks such as medical text summarization typically demands substantial computational resources. Parameter-efficient fine-tuning methods offer a promising alternative by updating only a small fraction of model parameters while maintaining competitive performance. This paper presents a comprehensive comparison of three adaptation approaches—Low-Rank Adaptation (LoRA), Prompt Tuning, and Full Fine-Tuning—evaluated across the Flan-T5 model family on the PubMed medical summarization dataset. Our experiments reveal that LoRA consistently outperforms full fine-tuning across all model scales, achieving 43.67 ROUGE-1 on Flan-T5-Large with only 0.6% trainable parameters compared to 40.82 ROUGE-1 for full fine-tuning. This finding suggests that the low-rank constraint provides beneficial regularization for domain adaptation. We analyze the performance-efficiency trade-offs in detail and provide practical recommendations for deploying medical summarization systems under varying resource constraints.
Parameter-efficient fine-tuning, Medical text summarization, Low-rank adaptation, Prompt tuning, Large language models
1 Anita R , 2 Himanshu Nainwal , 3 Koshti Vanshika Shaileshbhai , 1 Department of Computing Technologies, Faculty of Engineering and Technology, SRM Institute of Science and Technology,Kattankulathur, 603203, Chennai, Tamil Nadu, India, 2 Department of Computing Technologies, SRM Institute of Science and Technology, Kattankulathur, 603203, Chennai, Tamil Nadu, India , 3 Department of Computing Technologies, SRM Institute of Science and Technology, Kattankulathur, 603203, Chennai, Tamil Nadu, India
Retrieval-Augmented Generation (RAG) has been an effective tool in making natural language systems more relevant and factually accurate by providing more contextual information within their answers. However, in low resource languages such as Hindi and dialects such as Awadhi, the quality of the retrieved information is a big challenge due to differences in vocabulary, non-standard spellings, and the limited availability of curated data. In this study, we introduce a comprehensive hybrid dense-sparse Retrieval-Augmented framework for Hindi and dialectal Hindi (Awadhi), primarily aimed at enhancing the retrieval component. The suggested system combines many retrieval algorithms, such as BM25 for sparse lexical retrieval and multilingual sentence embeddings for dense semantic retrieval, to accurately capture both exact keyword matches and semantic similarity. FAISS indexes dense representations so that similarity searches can be done quickly and on a large scale. Retrieved document contexts are aggregated and structured to support retrieval-grounded response construction and downstream answer generation. Experimental evaluation demonstrates that hybrid dense–sparse retrieval significantly improves document relevance and ranking quality compared to single-method retrieval approaches. This study lays a strong foundation for Hindi and Awadhi RAG pipelines and sets the stage for the future inclusion of generative language models.
Dense Semantic Retrieval, Sparse Retrieval (BM25), Hybrid Retrieval Methods, FAISS-Based Vector Search, Low-Resource Languages
1 Jiawei Zhang, 2 Xin Zhang, 3 Xinyin Miao , 1 Senior Investment Analyst, PRA Group (Nasdaq: PRAA), Norfolk, Virginia, USA, 2 Data Scientist, PRA Group (Nasdaq: PRAA), Norfolk, Virginia, USA , 2 Data Scientist, PRA Group (Nasdaq: PRAA), Norfolk, Virginia, USA , 3 Senior Data Analyst, American Airlines Group Inc (Nasdaq: AAL), Dallas, Texas,USA
This paper utilizes four cutting edge deep learning architectures, namely VGG19, Xception, InceptionV3, and ResNet50, with transfer learning, image augmentation and two layers of regularization to be able to accurately predict the Alzheimers Disease classes under 33,982 MRI images with a 0.9563 accuracy, 0.9972 roc_auc, and 0.9559 f1 score in the testing scenario. By investigating the internal neural network structures and comparing the prediction performance, it provides the insight of how various deep learning architectures work differently under same conditions as well as the power of transfer learning and image augmentation in image-based classification and clinical diagnosis.
Deep learning, Transfer Learning, Image Classification, Neural Network Architecture, Regularization, Augmentation
1 Amy Tang, 2 Vinh Bui, 1 Buckingham Browne & Nichols School, 80 Gerrys Landing Rd, Cambridge, MA 02139, 2 California State University, Fullerton, 800 N State College Blvd, Fullerton, CA 92831
Overall, the app has 3 main features: authentication, stock analysis, and fetching trends with AI. Authentication verifies users’ emails and passwords and stores their information in the system using Firebase Authentication [1]. Stock analysis utilizes finnhub api to retrieve the stock data and display it to the user. The fetching trend’s function asks AI to search for trends and break it down using natural language processing to provide us with an analysis on these trends that affect the stock market [2]. The proposed application aims to analyze the influence of social media sentiment on the stock market through the use of artificial intelligence. Though its ability to use technical and quantitative analysis is limited, it still promotes financial literacy for beginners of the stock market.
Stock Market Analysis, Artificial Intelligence, Natural Language Processing, Financial Technology
1 Larry Yuheng Li, 2 Tyler Boulom, 1 Germantown Friends School, 31 W Coulter St, Philadelphia, PA 19144, 2 Woodbury University, 7500 N Glenoaks Blvd, Burbank, CA 91504
This project addresses the global decline in adolescent access to arts education by proposing MelodyCanvas, a virtual reality platform that enables students to share artwork, view galleries, and engage with a creative community. The system integrates Firebase authentication, AI-based content validation, and immersive gallery environments to create an accessible and interactive artistic space. Core challenges included building reliable 3D environments, ensuring accurate originality checks, and maintaining acceptable load times when retrieving artwork from cloud storage. These challenges were examined through experiments measuring AI classification accuracy and gallery performance. Findings revealed strong baseline functionality but identified the need for improved plagiarism detection methods and optimized loading pipelines. Compared with existing methodologies in vision analysis and VR interface design, MelodyCanvas offers a hybrid approach that balances technical feasibility with user experience. Ultimately, the platform demonstrates a promising way to support artistic development and cultural connection through accessible digital tools.
Virtual Reality, Arts Education, Creative Communities, AI Content Validation
1 Vimaladevi M, 2 Thangamani R, 3 Jayadharshini P , 4 Varshini B , 1 Department of Artificial Intelligence and Data Science, Kongu Engineering College, Perundurai, Erode , 2 Department of Artificial Intelligence and Data Science, Kongu Engineering College, Perundurai, Erode, 3 Department of Artificial Intelligence and Data Science, Kongu Engineering College, Perundurai, Erode 4 Department of Artificial Intelligence and Data Science, Kongu Engineering College, Perundurai, Erode, 5 Department of Artificial Intelligence and Data Science, Kongu Engineering College, Perundurai, Erode, 6 Department of Artificial Intelligence and Data Science, Kongu Engineering College, Perundurai, Erode
Agriculture in Tamil Nadu still largely depends on traditional methods, which leads to poor soil management and inappropriate crop selection. This article introduces RootLogic, a bilingual (Tamil, English) hybrid ensemble-based decision sup- port system that can identify soil nutrient deficiencies and suggest the best crops for a precision agriculture setup. The model under consideration combines Random Forest, AdaBoost, and Support Vector Machine (SVM) classifiers to produce better predictive accuracy and model generalization. It makes use of normalized soil and environmental attributes such as temperature, humidity, pH, rainfall, and macronutrients (N, P, K) for nutrient analysis and crop prediction at the same time. The experimental results on the combined agricultural datasets have shown an average accuracy of 99.55%, which is much higher than the performance of individual classifiers. The bilingual web interface makes it possible for rural farmers to have easy access to it so that they can be part of digital agriculture as well. The findings reveal that AI-based decision support systems like RootLogic can facilitate sustainable, data-centric, and location-specific precision farming, leading to enhanced productivity and better utilization of natural resources.
Precision Agriculture, Machine Learning, Soil Nutrient Deficiency, Crop Recommendation, Sustainable Farming.
1 N. Nalini, 2 Dadi Devaki Dhana Sree, 2 Chelika Lohith Paul, 2 Bolla Vijendra, 1 Assistant Professor, Department of CSE, School of Computing, Mohan Babu University, Tirupati, A.P, India, 2 UG Scholar, Mohan Babu University, Tirupati, A.P, India
Diagnosis and monitoring of Parkinson’s Disease (PD) traditionally rely on subjective clinical evaluation. This research presents a computer vision-based approach for analyzing finger- tapping videos to objectively measure bradykinesia. Motion features such as speed, amplitude, and rhythm are extracted and analysed using machine learning techniques. The system achieved 91% sensitivity, 97% specificity, and strong correlation (R = 0.740, p < 0.001) with clinical ratings. The proposed system provides an objective, contactless, and scalable solution for Parkinson’s Disease diagnosis and monitoring. It enables accurate clinical assessment and supports remote healthcare applications.
Parkinson disease, Bradykinesia, Computer vision, Machine learning, Pose estimation
1 Chaiho Wang, 2 Tyler Boulom, 1 Crean Lutheran High School, United States of America 2 Data Scientist, PRA Group (Nasdaq: PRAA), Norfolk, Virginia, USA , 2 Woodbury University,United States of America
Today around the world, agriculture faces many challenges like pest outbreaks, plant disease, inefficient resource use, and limited access to affordable monitoring tools for small to medium scale farmers. Agriculture faces persistent challenges such as pest outbreaks, plant diseases, inefficient resource use, and limited access to affordable monitoring tools for small- and medium-scale farmers. Without effective detection and intervention, these issues can result in decreased yields, financial losses, and long-term soil degradation. This project presents a smart farm robot that integrates robotics, advanced sensors, and artificial intelligence to provide real-time crop and soil monitoring. The system is built on a Hiwonder robot base with a Raspberry Pi, and leverages Gemini AI and OpenAI for plant image analysis, Firebase for backend storage, and a mobile application for farmer alerts. Several limitations emerged during development, including inconsistent image recognition under variable lighting and reduced navigation accuracy on damp or uneven terrain. Experimental testing confirmed that lighting conditions significantly impact AI performance, while soil type affects movement precision. Proposed solutions include adaptive preprocessing, LED-based lighting, and terrain-aware navigation controls. By addressing these challenges, the farm robot demonstrates potential as a scalable, low-cost precision agriculture tool. It offers a sustainable alternative to traditional monitoring methods, enabling farmers to make proactive, data-driven decisions that improve efficiency, reduce losses, and support long-term agricultural resilience.
Precision agriculture, Smart farming, AI crop monitoring, Agricultural robotics
1 Laura L. Zimny, 2 Andy Liang, 1 Northwood High School, 4515 Portola Parkway, Irvine, CA 92620, 2 California State University Los Angeles, 5151 State University Dr, Los Angeles, CA 90032
This paper addresses the technical challenge of bridging the gap between unstructured product data and personalized skincare guidance. The proposed solution, GlowLab, is an integrated mobile platform that synthesizes diverse data streams to optimize skincare routines efficiently and safely. The system is engineered around three core components: a persistent Data Service for tracking longitudinal skin health, a computer-vision enabled Product Analysis System that parses ingredient lists via the OpenAI API, and a Recommendation Engine that dynamically cross-references these inputs against user sensitivities [1]. Unlike systems relying on invasive biometric surveillance, this project addresses privacy and design challenges by utilizing secure local storage and user-reported data. Technical testing of the image recognition module demonstrated a 100% success rate across 20 test cases, confirming the systems ability to correctly parse complex labels into structured data [2]. The result is a robust, responsive application that transforms static ingredient data into dynamic, actionable health insights, providing immediate value for everyday consumers seeking safe, personalized guidance.
Skincare AI, Ingredient Analysis, Personalization, Mobile App
1 Jiaheng Su, 2 Marisabel Chang, 1 Viewpoint School, 23620 Mulholland Highway, Calabasas, California 91302, 2 California State Polytechnic University, Pomona, CA 91768
Parkinson’s disease diagnosis traditionally relies on subjective clinical evaluation and expensive medical equipment, resulting in prolonged wait times, substantial costs, and misdiagnosis affecting nearly 20% of cases. Many machine learning approaches require data from medical-grade imaging systems such as MRI, limiting accessibility. This paper presents a lightweight screening system utilizing facial movement analysis from standard video recordings to provide objective, accessible PD detection. The methodology processes videos through MediaPipe to generate facial mesh representations, extracting landmarks that are transformed into Action Unit features including eye aspect ratio, mouth aspect ratio, angles, velocity, and acceleration. A supervised Gradient Boosting classifier processes these features to distinguish PD patients from healthy controls. Experimental evaluation demonstrates 86.7% classification accuracy, substantially outperforming unsupervised K-Means clustering (46.7%). The proposed multi-region, dynamics-aware approach offers practical preliminary screening suitable for resource-limited clinical settings where specialist access remains constrained.
Parkinson’s Disease, Facial Action Units, FaceMesh, Computer Vision, EAR and MAR
1 Zihan Zeng, 2 Rodrigo Onate, 1 St. Michaels University School, Canada 2 2California State University, United States of America
Financial literacy remains critically low among young adults, with traditional education methods failing to engage digital-native learners. This paper presents the William Finance Group App, a cross-platform mobile application that addresses this gap through integrated real-time market data, AI-powered educational content, and community engagement features. Built using Flutter, Firebase, and OpenAI’s GPT-4, the application delivers personalized financial education through livestock data visualization, on-demand topic explanations, and gamified competitions. Key technical challenges including API rate limiting, LLM content accuracy, and real-time database synchronization were addressed through request queuing, structured prompting, and Firebase’s conflict resolution mechanisms. Experimental evaluation demonstrated high educational content quality (4.5/5 mean rating) and reliable API performance (99.2% success rate, 245ms average latency). The application synthesizes theoretical frameworks from digital financial literacy research into practical implementation, offering an effective solution for improving financial literacy among students through technology-mediated education.
Financial Literacy, Mobile Application Development, Flutter Framework, Large Language Models, Real-time Data Integration
Copyright © NLTM 2026