I spent six months teaching an AI to interview engineers. Here's everything it kept getting wrong

Most AI interviewers are built for an American in a quiet room. I found out the hard way that almost nothing about that holds up here.

For nine years I built backend systems. Distributed infra, fintech, a stint making autonomous-driving signal tests pass at scale. I am, by temperament, a person who trusts a well-defined spec.

So when I started building an AI that conducts voice interviews, I assumed the hard part was the AI. It wasn't. The hard part was that every off-the-shelf assumption about how an interview sounds was quietly wrong — and wrong in a way that punished the exact people I was building for.

Here's the field report. No pitch. Just the things that broke, and what we did about them.

1. It couldn't actually hear an Indian accent.

The uncomfortable open secret of voice AI: the speech models everyone builds on were trained mostly on American-accented English. Native speakers get word-error rates of 2–8%. Indian English routinely lands at 30–50%. That's not a rounding error — that's the model confidently mishearing every third or fourth word, then "evaluating" an answer it never correctly transcribed.

You cannot fix a foundation by being nicer to it. We had to build a transcription layer that expects Indian English as the default, not the exception. The accent was never the problem. The model's idea of "normal" was.

2. It treated thinking as failure.

Most AI interviewers cut you off after about a second of silence. Next question. But a pause isn't dead air — it's a person thinking, often in their second language, assembling a sentence before they commit to it. The tool was penalizing the most honest moment in the whole interview.

So we built what we call a five-tier silence ladder: the system distinguishes between gathering a thought, trailing off, genuinely stuck, and done, and responds to each differently. We gave it roughly 25% more pause budget than the Western default. The result is an interviewer with patience — one that waits the way a good human interviewer waits.

3. It had no memory, so it couldn't catch you.

A one-shot Q&A bot asks a question, scores the answer, forgets it, moves on. But a real interview is a narrative. The follow-up is where the truth lives. If you said your system handled 10,000 requests a second in minute three, and then described an architecture that couldn't do that in minute forty, a good interviewer notices.

We built memory across the full two-hour session and contradiction detection on top of it. We call it the honest mirror. It's not there to trap anyone — it's there because the gap between what we claim and what we can defend is exactly the thing interviews exist to find.

4. "You did great!" is not feedback.

The thing that surprised me most: encouragement is the enemy of improvement. Tools that end with a cheerful score teach you nothing. So every session ends with an honest report — filler words counted, pacing flagged, your single weakest answer rewritten the way it should have been said. Not a number to feel good about. A mirror to get better in front of.

The most common reaction we get isn't "great score." It's: "I didn't realize I said 'um' that many times." That sentence is the entire product.

5. The hardest part wasn't technical at all.

It was deciding what not to build. There's a whole category of "interview copilot" tools now that whisper answers into your ear during the live interview. It's framed as an edge. It's cheating, and everyone involved knows it.

I made an early decision: we will never be that. If you need a copilot to pass the interview, you'll need one to do the job — and that's a debt that comes due. We're a practice room, not a cheat sheet. Interviews are meant to be earned, not gamed.

That's six months of work in five problems. None of them were the problems I expected when I opened the editor at 4am back in winter. The AI was the easy part. Building something that actually listens — to the pause, to the accent, to the contradiction, to the person — that was the work.

If you've ever frozen in an interview when you actually knew the answer, you already understand the real problem we're solving. I'd genuinely like to hear about it in the comments.

Building @intervuesclub in public. Founding cohort opens July 2026.