Our voice interview platform
Your Next Interviewer Is a Link
Send an interview the way you send a calendar invite: whoever opens the link talks to a voice agent that listens, asks follow-up questions, and politely keeps the conversation on topic. In German or English, in the browser, nothing to install.
The most important decision sits under the hood: the language model carries the conversation, but which question comes next is decided by our server. We tried it without that oversight first — and learned why it is needed.
- Architecture
- the model speaks, the server decides
- Audio
- never touches our servers
- Languages
- German and English
Why the Server Has the Last Word
The first attempt was the obvious one: hand the realtime model the full question script and let it run. The result sounded good and still went off the rails — questions were skipped, reshuffled, improvised. For a structured interview, “sounds good” is not enough.
So today a supervisor sits on the server: it knows the interview guide, assesses where the conversation stands, and gives the model only the next step. The voice stays natural; the sequence stays binding.
The data path is a deliberate choice, too: the audio connection runs over WebRTC directly between the browser and the voice provider, using short-lived access tokens issued server-side. Our servers never see the audio. The interview goes out as a signed link, an open link, or a collective link for multiple sessions.
Areas
More from the AI Spotlight
Back to overview