AI experts often do not appreciate the constraints of embedded systems, and that the process of shoe-horning a conventional AI model into a tiny device can be overly complex and frustrating. We describe constraints that are typical to embedded implementations. We address these constraints by reducing the complexity of the deep net inference engine. We achieve that by minimizing the intra-network connectivity, eliminating the need for floating-point data, and replacing the multiply-accumulate operation with just accumulation. The resultant small-footprint, low-latency deep nets are suitable for embedded applications that employ any 8/16/32/64-bit MCU, DSP, and FPGA. This solution is targeted at IoT smart sensors for inertial, vibration, temperature, flow, electrical, and biochemical measurements in battery-powered endpoints. Applications include healthcare/industrial wearables, robots, and automotive.