Deploying efficient machine learning models, particularly deep neural networks, is critical for edge-ready AI applications, e.g., for detecting persons in office rooms, which is a relevant task in building automation. These models aim to achieve a trade-off between complexity and performance. We apply and evaluate three key approaches to achieve this balance: neural network architecture search, post-training quantization methods, and mixed precision training for deep neural network models. In addition, we compare the outcome of these methods with classical machine learning techniques, such as Support Vector Machines and Random Forest. The results demonstrate that mixed-precision training balances model accuracy and complexity well. In contrast, post-training quantization methods significantly reduce model complexity but at the cost of decreased model accuracy.