This paper presents an edge neural architecture search method for human presence detection using Wi-Fi signals. The proposed method, WiFiNAS, is a derivative-free, cell-based technique to design tiny 1D convolutional neural networks for microcontrollers. The process iteratively explores topologies by adjusting the number of convolutional filters and computational layers, based on validation accuracy under micro controller memory constraints. This deployment happens by the unified core technology, which imports and profiles memory usage, operations, and latency on tiny microcontrollers. This information provides feedback to the search process. The final 8bits model achieved a classification accuracy of 98.57%. Real-time inference was achieved on various STM32H7, STM32MP1, and STM32MP2 with neural processing acceleration in 78.6 ms. This work ensures full privacy by keeping data local, within a home environment.