It should be based off all three platforms. But to start off you should create a standalone AI assistant, then refactor to each of the 'already available platforms based on feedback.' The latter should be easier to scratch if you have your own backend to digest and reply via api layer from.