Summary 1 highlights an innovative approach to inspecting transmission lines using AI, where advanced computer vision and Deep Learning techniques are implemented. The study demonstrates how these technologies can detect faults in transmission lines, reducing the need for manual inspections. By training high-accuracy models on powerful hardware and optimizing them for deployment on devices like the Nvidia Jetson Nano, the system effectively analyzes images from drones to identify critical issues such as broken conductors. Summary 2 deals with the challenges in AI training services, particularly regarding correctness and resistance to attacks like data poisoning during training. It introduces a method that addresses GPU nondeterminism through higher precision training and adaptive techniques, allowing for exact replication at FP32 precision on different NVIDIA GPUs. The approach aims to improve efficiency in training models like ResNet-50 and GPT-2, reducing storage and time costs compared to traditional methods. Summary 3 presents advancements in optimizing single-GPU usage for large AI models. By utilizing arithmetic-intensity-guided scheduling, the AIGPU model enhances kernel execution overlap, formulated as an integer linear programming problem to optimize performance. Testing on NVIDIA RTX 3080Ti demonstrates significant improvements in throughput and GPU utilization compared to existing methods, showcasing a more efficient approach to maximizing single-GPU capabilities. Summary 7 focuses on the development of an AI algorithm to distinguish between two types of myeloproliferative neoplasms, prePMF and ET, using bone marrow biopsy whole slide images. The RetCCL neural network, used in the model, achieved high diagnostic accuracy and specificity, with evaluations performed efficiently on consumer hardware. This solution aids in identifying patients for specific treatments or trials, offering a reliable and cost-effective diagnostic tool.