IoT device development has traditionally relied on a monolithic approach, with all firmware developed by a single vendor using a rigid waterfall model, typically in C, and infrequently updated. This paradigm is insufficient. For AI-enabled IoT devices to reach their potential, developers must be able to easily program and update low-cost, low-power AI-capable sensors, such as Sony’s IMX500. In this talk, we’ll present Wedge—a combination of runtime, device agent and cloud service. Wedge automates software life-cycle management for devices, provides isolation and enables agile development. Wedge is based on WebAssembly, a binary instruction format for a stack-based virtual machine. Wedge’s runtime and memory overhead are within 2x of native C/C++ code, and it supports languages such as Python and AssemblyScript. We’ll also present the vision sensing pipeline, a sensor data processing layer we have developed on top of Wedge, which is accessible via REST API and a visual interface.