<@UMWLCTHJ7> I don't believe in private discussion...

# science

[altavir]

@Amir Gur I don't believe in private discussions on non-private topics, so I will answer here. The question was about implementing kotlin wrapper for for TensorFlow. The answer consists of few statements:
• There are already official java wrapper, which could be used from kotlin without effort: https://www.tensorflow.org/install/lang_java
• Currently the question is not about the wrapper, but about developing the API which could use kotlin advantages to the maximum. For that we need to understand what features of TensorFlow are really needed. I've created a repository for such proposals here: https://github.com/mipt-npm/KEEP-math, we can use it to formulate the requirements.
• As soon as API is ready, we can use it both for TensorFlow and to devise our own pure-JVM implementation which could be used as a fallback when deployment simplicity is more important than performance. This is basically the idea behind Kmath: create universal interface with simple implementations and provide a mechanism to use high-performance implementations instead.