All the floppy drives I've ever seen had a little mechanical arm attached to a microswitch that detected the presence or absence of the write protect notch. VHS cassettes worked similarly.
I've had great success doing transcription on-prem with Speechmatics: https://www.speechmatics.com They offer both a VM and a Docker image that you can run behind your firewall, and even license offline if you really need to. I use it for generating closed captions of videos, but you could build a transcription tool out of it as well. Their engine's accuracy is the best I've found out there, too, which is a nice bonus.
The exact opposite for me -- I’ve been interested in using FastAPI with SQLAlchemy for some time, so this project is exactly what I’ve been looking for.
Also of note: HLS supports byte range addressing, so you can create the various stream files as single .ts files rather than a collection of segments per stream. A client can use the Range HTTP header to select the window of bytes it wants for the stream/bandwidth slot it wants. This mode is supported from something like iOS 5 and up, and ffmpeg has flags to produce such streams.
I'm curious to learn how API response caching is affected by GraphQL. In a REST setup, there's a possibility that API responses can be cached. But if the response structure is dictated by the client, it seems like responses might differ and not be cacheable.
I imagine it could work like Android 9-patch PNG drawables, which have an extra ring of pixels around the perimeter that mark content zones and stretching zones.