Developer Replaced LLM-Driven Interview Flow with a Finite State Machine for Reliability

A developer building an AI-powered technical interviewer found that relying solely on a large language model to manage interview flow caused consistent failures, including premature scoring, repeated questions, and accidental answer giveaways. The core problem was structural: LLMs lack reliable state-tracking across conversation turns, making them unsuitable as sole controllers of multi-stage workflows. To fix this, the developer implemented a finite state machine that owns all phase transitions, such as moving from coding to evaluation only after code is actually submitted and run in a sandbox. The LLM was retained for content generation within each state — writing questions, hints, and scorecards — but stripped of any authority to decide what happens next. The result was a more reliable and, counterintuitively, more naturally conversational product, since the LLM could be expressive within tightly controlled boundaries.