It's not impossible. But there is a very real gap between "this is cool in a research paper" and "this is deep learning that works in real life".

It's a large gap, covering everything from application topics, to data quality, to the need to actually run the damn think in a production setting with scalability, availability, error handling, etc.

Production applications of deep learning aren't particularly glamorous, they're not the "next big thing" right now. Rather they're improvements of existing applications.

Google's on device live captioning works really well, but still somewhat niche, and requires special / higher end SoC's to run.