The author teamed up with an AI assistant to build a pipeline that puts the Ryzen AI NPU in their AMD 8845 (Hawk Point/PHX) to work, running five vision models locally without cloud or discrete GPU. The pipeline executes YOLOv8m (object detection), ResNet-50 (classification), RetinaFace (face detection), MobileNetV2 (classification), and RCAN 2× (super‑resolution) using INT8 inference on the NPU. Although the NPU currently only exposes CNN INT8 capabilities, the author points out that LLMs would need BF16 support, indicating a present limitation and a path for future expansion.