I gave my hiring copilot a voice, because reading practice questions off a screen never simulates a real interview
For a while, HiredCopilot's interview prep was text: here's a likely question, here's how to answer it. It worked, but it never felt like an interview. Reading a question off a screen and typing a response doesn't create any of the pressure of someone actually asking you something and waiting for your answer.
So I rebuilt it as a real spoken interview. You talk to an AI interviewer over an actual voice call, grounded in your real resume and the specific job you're prepping for, not generic questions. It can end the call itself once it's actually done, the same way a real interviewer would wrap up, instead of running until a timer says stop.
Getting there wasn't clean. Three separate production bugs, each only showable by actually running a real call: a denied microphone could silently burn someone's one free trial, a client library's own connection function returned a value that looked like a promise but wasn't (so a failed connection just silently pretended to succeed), and I had to find that out the hard way before the retry logic could ever work right.
It replaced the old text-based tool entirely rather than living alongside it. Free to try, paid tiers get more minutes. Would genuinely like to know if a spoken practice round changes how prepared people feel walking into the real thing.
The strongest part is replacing text prep with an actual spoken interview rather than simply adding voice as another feature. The production bugs also show why the shift to real calls changes the product meaningfully.