Why I built an AI interview copilot, and what surprised me technically

I kept watching strong engineers bomb interviews. Not because they couldn't do the work, but because the format is brutal: a stranger watching, a clock running, and your brain choosing that exact moment to go blank.

So I started building LastRound AI for myself. It began as a real-time copilot for live interviews and grew into a fuller prep stack, voice mock interviews that score how you actually respond, an ATS resume builder, and cited company-by-company interview breakdowns.

A few things that surprised me building it:

The hard part wasn't the AI, it was latency. A suggestion that lands 8 seconds late is useless mid-conversation, so streaming speech-to-text plus a rolling context window mattered more than any model choice.

A cheap "is this even a question?" classifier before the expensive model saved more cost and latency than any prompt tweak.

People don't actually want more LeetCode. They want to stop freezing.

Still early and figuring out distribution. If you've shipped anything real-time + AI, I'd genuinely love to hear how you handled the latency vs quality tradeoff.