Gongcheng Kexue Yu Jishu/Advanced Engineering Science

The multimodal approach uses the heterogeneous sources of data to promote perception, inference, and decision-making in intelligent systems. A multimodal framework does not just use one channel, like text, audio or facial appearance, but combines the corresponding streams of information and reaches more reliable and context-sensitive predictions. The present study is a multi-modal emotion recognition-based and age filtering-based developed advanced music recommendation system, which incorporates real-time multimodal emotion recognition and age filtering to deliver very personalized song recommendations by Indian/Bollywood genre. The system uses a hybrid machine learning design with an integration of Face-API.js to analyze facial expression with up to 87 percent recognition accuracy and a DistilRoBERTa transformer-based sentiment classifier to analyze text-based inputs with 85-90 percent accuracy. In addition to providing a more personalized approach to music, a selected dataset of 600 Bollywood music tracks is enriched with psychoacoustic data (valence, energy, tempo, and danceability) and mapped to affect-to-audio features with a weighted scoring model. It allows narrowing the gap between perceived emotional states and musical qualities, thus making the recommendation more relevant in the dynamic real-world context. Experimental measurements affirm good performance of the system with end-to-end response times held down to less than 2 seconds and face-processing throughput of 2530 FPS. While doing the Practical implementation a user study of around 50 participants was done in two weeks with a mean satisfaction score of 8.5/10, an 83 percent recommendation intention rate, and 72 percent repeat engagement. On the whole, it can be concluded that hybrid multimodal emotion analysis, demographic adaptation, and cultural relevance are important to further the development of next-generation music recommender systems that can provide reliable, affect-sensitive, and human-centered interaction experiences.

Over time, customer demands and service requirements evolve, making it essential for service industries to adapt to new technologies. This underscores the importance of upgrading existing systems and processes. This study focused on evaluating the current inventory and monitoring system for medical supplies in various municipalities and developing an improved system that delivers more accurate and efficient results. Through thorough analysis, the researcher designed a user-friendly, efficient, and more accurate system. The recording and encoding of beneficiary information have become paperless and faster compared to the old system. The study not only enhanced the system but also improved the services provided by rural health centers.

Cotton is one of the most important cash crops in the world, and many plant diseases have a big effect on how much it grows. To protect food and economic security, it is important to find and treat cotton plant diseases as soon as possible. Machine learning (ML) and deep learning (DL) methods have been widely used in the last few years to automatically find plant diseases using pictures of leaves. Nonetheless, current studies present findings derived from varying experimental conditions, datasets, and evaluation metrics, complicating direct comparisons. The agricultural productivity and the textile industry are at a high risk due to cotton plant diseases. These diseases should be detected early and when it is necessary as this will make the loss minimal and may lead to better yield in crops. This paper will involve a detailed comparative performance evaluation of deep learning and conventional machine learning algorithms in the detection of cotton plant diseases. Deep learning networks, such as ResNet18 and Swin Transformer, are tested in comparison with the traditional machine learning algorithms, such as Logistic Regression, Random Forest, K-Nearest Neighbors (KNN), and Gaussian Naive Bayes. The experiment is done on a publicly available dataset of a cotton disease. Findings indicate that Transformer-based models are much better than CNN and traditional methods because they can capture global contextual relationships. The Swin Transformer has the most accurate and ROC-AUC and is thus suitable when faced with complex plant disease classification problems.

While developing Generative and Training Language Models, there is a problem called "hallucinations", which refers to generative language models producing responses that seem reasonable given a context but are actually factually incorrect. In this paper, we present a new Optimised Retrieval Augmented Generation Pipeline that seeks to eliminate the hallucination problem, by dynamically grounding responses to external, retrievable knowledge. We have developed a system that leverages the following: (1) LangChain for orchestration; (2) LangGraph for workflow management; (3) Crew AI for tier-3 agent-to- agent coordination; (4) ChromaDB, a vector data base; and (5) the Gemini API for Generative outputs. Key architectural distinctions include: Intelligent Semantic Chunking; Adaptive Retrieval Strategies; and Structured Multi-Stage Prompts, and we back those distinctions up with a rigorous evaluation framework. Our experimental evaluation demonstrates a statistically significant improvement over our baseline LLMs, with a factual accuracy score of 93.4% compared to 76.2% for the baseline; a decrease in identification of "hallucinations" from 18.3% of all responses to under 4.1% of all responses; and a citation accuracy score of 96.2% of all responses produced by this system in the shortest time frame (3.4 seconds). The modular architecture of this system will allow for effortless adaptation across different domains, by simply updating the knowledge base, without having to retrain any models. This work establishes a practical scalable framework for safely deploying factually grounded, trustworthy generative AI models, and provides an example of how organisations with limited resources in machine learning can easily access sophisticated RAG capabilities.

In critical situations such as medical emergencies, domestic violence, or fire incidents, people, particularly those with speech or physical impairments, may not be able to seek help by conventional means. This work presents a gesture-based real-time emergency alert system that uses AI-driven technologies integrated within an edge-cloud architecture. The system runs on a lightweight edge device, such as a Raspberry Pi, and captures live video to detect predefined emergency gestures using a computer vision-based approach with landmark extraction and gesture classification. Upon confident detection, the system triggers automated alerts through SMS or voice calls using a cloud back-end hosted on AWS. To improve reliability, it incorporates confidence smoothing, gesture conflict resolution, and filtering mechanisms to reduce false positives. The proposed framework offers a scalable, low-cost and accessible enhancement to existing emergency response systems.