Modern AI systems demand rigorous approaches to optimization, efficiency, and reliability. This course teaches students how to systematically improve the performance of contemporary large AI models by integrating principles from machine learning, experimental design, and statistical optimization. Students will learn foundational and applied concepts including network architectures, hyperparameter tuning, ablation analysis, model distillation, mixture-of-experts (MoE), and fine-tuning strategies. Building on classical statistical tools such as fractional factorial designs and response surface methodology (Montgomery), students will design, execute, and analyze controlled experiments to optimize neural network performance while minimizing computational complexity. The course emphasizes hands-on skills through practical labs and a capstone project in which students apply DOE methods to tune a neural network, ablate ineffective components, and distill the model into an efficient student model. By the end of the course, students will be able to systematically evaluate model architectures, identify drivers of performance, and reduce compute and infrastructure costs without loss of accuracy.