There are some situations where character recognition is easy enough that it should be possible without a ton of machine learning, so I am creating a JS/wasm client-side library that can accurately identify characters from pen or touch strokes. For some reasonably accurate results for the English alphabet, it currently zips down to about 50KB all-inclusive and takes just a few ms to identify a character on most devices.
I have tried to make it easy to train your own character set if anyone wishes to try that out. It is still in the early stages but I am hopeful it will eventually be good enough for many use cases.