In many use cases of smartphones, such as activity recognition, we need machine learning models to solve the underlying problem. However, machine learning models can grow vastly under certain circumstances and consume many computational resources. Thus, it is crucial to consider the resource factor in the model development. In my Diploma thesis, I found for human activity recognition models that many research papers primarily focus on model accuracy and lack deeper considerations of resource consumption. I developed a 3-step engineering process for machine learning models that uses resource-enhanced grid search to identify a model configuration with the best resource-accuracy-tradeoff. In this talk, I will outline my process and learning outcomes, and discuss why porting and engineering machine learning models for smartphones is still an ongoing problem.