<p>The rapid convergence of artificial intelligence and edge computing has created an urgent need for energy-efficient intelligence operating under strict computational and power constraints. As billions of devices move intelligence closer to the data source, modern edge systems must support real-time inference, guarantee privacy, and remain operational on milliwatt-level power budgets. This manuscript presents a comprehensive and updated review of low-power AI and signal processing techniques designed for resource-constrained embedded platforms, incorporating key developments from 2018 to 2025. The review synthesizes lightweight neural architectures, hardware–software co-design principles, quantized and pruned inference pipelines, TinyML frameworks, and low-complexity signal processing front-ends, in addition to examining neuromorphic and federated edge-learning trends. The quantitative dimension of the abstract has been reinforced by highlighting the comparative evaluation of latency, energy consumption, and accuracy across representative low-power models. The review extends prior surveys such as Gholami et al. (A survey of quantization methods for efficient neural network inference, 2021) by offering a broader cross-layer treatment that integrates signal processing, model optimization, and hardware evolution into a unified framework. Newly added graphical analyses illustrate the power–accuracy trade-offs in representative architectures, and a classification scheme is presented to organize edge-AI tools, frameworks, and hardware platforms. The manuscript concludes with an expanded discussion of emerging neuromorphic systems, federated micro-controller learning, event-driven sensing, and real-time embedded inference.</p>