Phone or Zoom — when an answer goes vague or rehearsed, the agent asks the next question, and the next, until the rubric criterion is actually answered. Every follow-up is logged in the transcript with the rubric line it tested.
Transcript · live
Trusted by recruiting teams shipping audit-grade hiring
01 · Two interview brains
Doubao mode runs Claude Sonnet through eight explicit tools (speak, ask_question, probe_deeper, …); the voice engine is just a mouth. OpenAI / Gemini mode is autonomous: the voice model has semantic VAD and runs the conversation, while Claude monitors as a safety supervisor only. Both modes share the same role-adaptive system prompt and rubric, so a recruiter never has to know which one ran the call.
02 · Tools
speak · ask_question · probe_deeper · move_to_qa · end_interview · get_meeting_status · wait_for_update · check_health. Every call lands in interview_audit.json with a timestamp and arguments. probe_deeper is the differentiator — when an answer goes thin the agent fires it against the specific rubric criterion that is still open, and keeps drilling until the answer covers what the rubric asks.
03 · Safety
A 45-minute hard cap (MAX_INTERVIEW_DURATION_SEC) ends every interview cleanly. A meeting observer polls the Lark API every few seconds — if the candidate leaves, the agent says goodbye and ends. A separate supervisor scans active interviews every five minutes and requeues any that look stuck. Farewell is the only forced speech in the whole pipeline.
Real-time
The voice model runs the conversation; the supervisor model watches the rubric. Both end the call within the 45-minute hard cap so no candidate ever floats.