The growing demand for privacy-preserving AI on edge devices makes Federated Learning (FL) an increasingly vital paradigm. This project aims to advance the discovery and deployment of optimal AI models in FL environments by developing and evaluating DeepFedNAS, a novel framework for Neural Architecture Search (NAS). Traditional NAS in FL faces significant challenges including high computational costs, communication overheads, and statistical heterogeneity across distributed clients. DeepFedNAS directly addresses these by enabling the efficient, principled, and hardware-aware discovery of high-performing neural architectures. Our goal is to rigorously test DeepFedNAS's ability to identify optimal models under diverse FL settings, ensuring robust performance and efficient resource utilization crucial for real-world, on-device AI applications. Motivation and Goal Federated Learning is critical for privacy-conscious on-device AI, allowing models to be trained on distributed data without sensitive information leaving local devices. However, manually designing efficient neural architectures for these diverse and resource-constrained edge environments is impractical, making NAS a necessity. Existing FL-NAS methods struggle with scale, often requiring extensive computation and communication, thus limiting their practical adoption. This project focuses on overcoming these barriers through DeepFedNAS. Our exploration will focus on critical aspects: Principled Architecture Discovery: Developing a methodology to guide the NAS process with mathematically-sound principles, moving beyond random or brute-force search. Resource-Efficient Supernet Training: Investigating strategies to train powerful supernets without prohibitive computational demands on clients or central servers. Rapid Hardware-Aware Deployment: Enabling the swift discovery of customized neural architectures that meet specific hardware constraints (e.g., MACs, parameters, latency) of edge devices, post-training. Robustness to Heterogeneity: Evaluating DeepFedNAS's performance and stability across varied data distributions (non-IID) and system capacities common in real-world FL. Our goal is to establish DeepFedNAS as a state-of-the-art solution for federated NAS, identifying configurations that best support efficient, high-performing, and privacy-preserving AI deployment in diverse on-device scenarios. Experiments will gauge its superiority in performance, efficiency, and adaptability compared to current approaches.