AI Interviewer

Transcription and answering are separated by a detection stage that decides what the last genuinely unanswered question was — which is not the last sentence, and often not phrased as a question at all. That stage is the difference between a useful assistant and a live transcript with noise attached. There's also a manual override box, because the honest design assumption is that detection will sometimes be wrong.

The window calls Windows' `SetWindowDisplayAffinity` with `WDA_EXCLUDEFROMCAPTURE`, so the compositor itself omits it from any capture stream. On Meet, Zoom, Teams or OBS it renders as a black rectangle while staying fully visible on the physical display. This is an OS-level guarantee rather than a heuristic that breaks when the conferencing app updates.

Whisper runs locally on the GPU — audio never leaves the machine, and local inference removes the network round-trip that would make transcription lag unusable. Only the detected question text goes out, to OpenRouter, which keeps the model swappable between Gemini Flash, Claude and GPT without touching the pipeline.

Generation is conditioned on three things: your parsed CV, a `tone.md` voice guide, and the pasted job description. Without the tone guide the output is recognisably model-shaped — balanced, hedged, slightly too fluent to be speech. The guide is what makes an answer sound like something a person would actually say out loud.