What we learned.
Generic speech AI is not enough. Names, places and casual speech broke off-the-shelf models, so we built a domain-tuned pipeline.
People say several things at once. One sentence often holds two or three requests, so Recho splits them into separate tasks.
Trust needs a screen. Showing the words before sending, with a re-record option, is what makes staff comfortable relying on it.
Hardware is worth it. A dedicated device beats a phone app for habit: no unlocking, no notifications, one button to speak.