This framework was written out of necessity. In one of our projects, an intelligent caching system, the model was too big and slow that was very expensive in hardware terms.
During development a radical idea arose to design the backpropagate algorithm in a very different and unusual way, a proof of concept was done and ... it worked better than expected

The design was made entirely in C to guarantee the maximum possible speed. It behaves like a regular perceptron. Internally optimizes calculations by refusing to update all weights in each learning cycle. We also developed python libraries to be able to use the PRModel API from this language.

The results are spectacular in both speed and memory usage. To compare the result of a standard classification problem against Keras using TensorFlow, the Google framework for ML. In this case the problem determine in which country a hotel is located, with a database of addresses and hotel data, as it is a big corpus of a bag different words the X vector has 79.810 features. On 15 epochs learnings:

This technology allows us to address problems that are beyond the reach of machile learning due to size and provides us with a very important competitive advantage. PRModel can be expanded to support other types of neural networks, such as LSTM or convolutional networks. PRModel will be even faster when GPU support is ready and can span infinite models regardless of available physical memory because internally it manages weights as b-tree, PRModel can be automaticaly be connected to a database like mysql.