The Peltarion Cortex platform, used in all of our projects is a development and production system that adds an industrial grade software engineering layer to the research oriented open source deep learning libraries.

Our Peltarion Synapse / Spine development environment and production server, have a proven track record of reliability and engineering quality. They have been in production use by reliability critical sectors such as the banking industry for years.

Peltarion Cortex builds on that heritage, extending it to the complexities of dealing with very large models trained on specialized hardware while retaining the easy of use, maintainability and documentability. 

The Cortex platform runs on Ubuntu Linux (GPU support) and on most unix based systems (including OSX) on CPU (typically too slow for training, but fast enough for using the system). It has full version control of blocks and models as well as states generated during training and data sets. Models are serialized with their definitions intact and all system configurations are logged and version controlled. This ensures reproducibility as well as traceability. Peltarion Cortex fully supports continuous training with automatic validation and continuous integration of new models.

Peltarion Cortex uses either Torch7 or TensorFlow as a computational backend. At this point we prefer Torch, but that may change as TensorFlow matures.