Neural Architecture Search (NAS) and extreme quantization tackle TinyML efficiency from complementary angles: NAS targets accuracy, full binarization aims at lowering the computational load of deep architectures. The recently introduced NAS‑BNN framework combines NAS with binarization to automate the discovery of high‑accuracy and low‑compute Binary Neural Networks (BNNs). Targeting person‑presence detection, we train a five‑stage supernet on 414 k images from the new WakeVision (WV) benchmark and perform an evolutionary search over 250 sub‑architectures. The resulting Pareto front spans 3–8 million operations (OPs). Our top model achieves 86.69% Top‑1 accuracy at 6.03 M OPs, surpassing the MobileNetV2 0.25 and MCU‑Net‑320 kB benchmarks while requiring 12–18 × less compute. Even the smallest network discovered (only 3.84 M OPs) achieves 79% accuracy. These findings demonstrate that NAS‑discovered BNNs can match or exceed state‑of‑the‑art accuracy while keeping a very low amount of operations, and establish WV as a realistic large‑scale benchmark for TinyML research.