Notes from building the agents.

A frontier real-time voice model isn't broken when it repeats a question — it's doing exactly what it was built to do. Here's why that's a problem for hiring specifically, and what holding it to a script actually requires.

Voice interview AI is almost always demoed on software engineers. We ran it on aerospace, agriculture, healthcare, and administrative roles on purpose — here's where a resume-anchored system actually breaks.

Keyword matching is easy to build and easy to game. Here's how the Interview Evaluator weighs substance and structure instead — and why it shows its reasoning, not just a number.