Volume 4 of 10 in the Engineering-to-Research Monograph Series. The default machine-learning pattern, train a large model in the cloud and send data to it, breaks at the edge, where data originates on power-, memory-, and bandwidth-constrained devices that must often decide in real time and in the physical world. This report treats edge AI as a four-layer co-design problem (sensing, model, runtime and hardware, software engineering) and argues that its central engineering act is principled reduction: shrinking a model while preserving both accuracy and the explainability a non-expert user needs to trust a prediction. It contributes a four-layer reference architecture and grounds it in two built systems: AgriEdge, a plant-disease classifier fusing thermal and visual imaging on an NVIDIA Jetson Nano (a MobileNet-based convolutional network reaching about 94.9 percent validation accuracy across 38 classes), and SensoryPi, a Raspberry Pi cyber-physical security system with on-device facial recognition over MQTT. It closes by connecting edge explainability to AI governance.The paper and figures are licensed CC BY 4.0; companion code is released under the MIT License. This work contains no confidential or proprietary employer information.