This thesis is about the design, implementation and evaluation of a framework that optimizes neural networks for mobile applications. With the use of neural networks in practice and on mobile devices, like the Hololens, a mixed reality headset build by Microsoft, better and smaller neural networks are in need. This framework measures the performance of neural network and creates two new networks based on a threshold. If the given network performs worse on a trainings instance than the defined threshold, it is marked to be used for a further training. These marks are used to train a deciding network and a supporting network. The supporting network is trained on the trainings instances that were worse than the threshold. The deciding network will be used for inference, when the new predict function is called, the deciding network will decide if the instance should be predicted with the main network or the supporting network. This approach is also enhanced with extra iterations, with new set of neural networks as the given network to be improved. Such that multiple deciding and supporting networks can be found iteratively. Finally, a network is trained, that combines all new deciding networks into a super decider which will predict for a given instance the corresponding main or supporting network. This framework is evaluated on two different datasets, and seems to improve the performance of a network trained on one of the dataset. The end result is a framework, that can be used with little configuration, to optimize neural networks, that perform bad on significant edge cases.