Citizen insecurity is one of the most important problems in our society. We have reviewed the investigations that predict crimes with different techniques. Due to this, in this research we propose an architecture de neural network capable of predicting the possible crimes that a person could commit, after answering a form about their family and social conditions. In the research, three architectures have been tested to look for the improved architecture of the neural network that predicts the crimes. We conducted the experiment using data from the 2016 Peruvian National Penitentiary Census, which consists of 14,000 records. We conclude that the neural network architecture with 24 input neurons, 21 neurons in the first hidden layer, 21 neurons in the second hidden layer and 19 neurons in the output layer is the most recommended for predicting crimes with the mean square error obtained was 0.278100. We have also implemented a web application as a tool for the institutions that administer and punish crimes.