The author grew tired of expensive and restrictive cloud AI services, which they deem a subscription trap. They built a pocket-sized, low-power device using a Hailo-8 AI accelerator and a Raspberry Pi 5 to perform AI tasks locally. This approach eliminates latency, cost, and privacy concerns associated with cloud-based inference. The Hailo-8 chip is praised for its power efficiency and performance in edge computing. The author details the hardware components, including the Raspberry Pi 5, Hailo-8, an NVMe base, and a touchscreen, costing around $220-$280. Building a functional and portable device required careful consideration of cooling and stable power delivery. The software stack involves a dataflow architecture, combining the Hailo-8 for heavy computation and the Raspberry Pi's CPU for sequential tasks. This local setup achieves significantly lower latency for vision-language tasks. The author emphasizes that moving inference locally is not just about performance but also about privacy and data control. The device offers robust capabilities like real-time object detection, pose estimation, and image embedding generation. The composability of local models allows for complex, chained operations without incurring API penalties. The author concludes that building such a device offers independence from cloud providers and a deeper understanding of AI deployment.