This paper introduces a method to boost diagnostic accuracy from blood tests, even when labelled data is scarce. By combining consistency regularization with pseudo-labels on unlabeled CBC reports, our approach makes the most of limited data. We also incorporate uncertainty-aware calibration, improving safety and reliability for clinical use. On anemia, deficiency, and infection tasks, our method consistently outperforms fully supervised models, delivering higher recall and AUC, and demonstrating real-world promise for safer, scalable clinical AI.

This project combines satellite and weather data with on-the-ground IoT sensors to deliver accurate yield forecasts and tailored crop recommendations. An interactive dashboard brings together multi-year market trends, enabling growers to plan with greater confidence, allocate resources more effectively, and make faster, data-driven decisions. By putting timely insights in farmers’ hands, this system aims to drive smarter, more resilient agriculture.

This project integrates property listings, points of interest, and transit accessibility data to deliver highly accurate real-estate valuations using GBM and stacked ensemble models. By applying SHAP for interpretability, it highlights key factors like neighborhood amenities and commute times behind each valuation. The result: pricing guidance that not only outperforms single-model approaches, but is also transparent and auditable for users and stakeholders.

This study evaluates both classical and ensemble machine learning models for early cardiovascular disease (CVD) risk prediction, with a strong focus on maximizing recall to minimize missed diagnoses. It emphasizes the importance of interpretable features, such as cholesterol levels, chest pain type, and ST depression, to ensure clinical transparency. The result is a practical framework for developing effective, easy-to-use screening tools, especially suited to resource-limited healthcare settings.

This mobile app bridges the gap between natural language and databases, converting plain-English questions into executable SQL using intent parsing and slot-filling. Designed with education in mind, it lowers the barrier for students learning databases and boosts task completion rates, as shown in user studies. The app runs offline with lightweight models, making it accessible and effective for classrooms with limited connectivity.

This system analyzes AIS (Automatic Identification System) vessel trajectories using speed and heading patterns to automatically detect suspicious activity near protected maritime zones. By flagging potential violations, it reduces the manual workload for enforcement teams and helps prioritize patrols where they matter most. The model is designed for adaptability, making it easy to deploy in new maritime regions with minimal additional training.

This project presents a fuzzy rule-based controller for autonomous vehicles that decides between braking and steering in complex urban traffic scenarios. Inspired by how humans make split-second decisions, factoring in stopping distance and obstacle proximity, the controller reduces hesitation and adapts to uncertainty. Simulations demonstrate that this approach leads to fewer near-misses compared to traditional fixed-logic systems, offering a safer and more adaptable solution for real-world driving.