Situation
I wanted to build something I would use rather than another exercise, and the small repetitive things I do at my desk seemed like a reasonable target. Opening apps, checking the time, putting music on.
Task
Build a working voice assistant in Python. Listen, transcribe, work out the intent, run the right thing.
Action
The pipeline listens for a wake phrase, transcribes what follows using speech recognition, and matches it against a set of command handlers. Each handler does one thing, so adding a new capability is writing a function and registering it rather than editing a parser.
Result
It works for the commands I built it for.
What I did not expect was where the difficulty actually was. I assumed the hard part would be recognition accuracy, but transcription was mostly fine. The hard part was intent. People phrase the same request several different ways, and when the assistant gets it wrong it needs a sensible way to recover rather than confidently doing something you did not ask for.
So most of my time went into the failure path rather than the happy path, which is not where I expected to spend it.