An AI voice agent that runs the first interview
For the recruiting platform Evidio, we built an interview agent that joins the call and runs the screening conversation itself – tailored to the role, with a score and briefing at the end instead of loose notes.
The same first interviews, over and over
In screening, most of the time goes into first interviews that largely follow the same path. The same questions, the same order, and loose notes at the end that are hard to compare between candidates.
There was no off-the-shelf tool with the depth needed – above all none with a voice agent that holds a real conversation instead of working through a form.
An agent that joins the call
When the call room is created, the role, interview dimensions and language are passed in. The agent loads the CV data and runs the interview tailored to the role – as a real conversation, not a form being filled in. Technically, the interviewer is one of several AI agents we build for specific workflows.
How an interview runs
Create the room
The role, interview dimensions and language are passed in when the call room is created. That tells the agent what matters in this particular conversation.
The agent runs the conversation
The agent joins the room, loads the CV data and runs the interview tailored to the role. Speech recognition and output run through Deepgram, the conversation itself through Claude.
Score & briefing
After the call, scores are produced along the interview dimensions plus a detailed briefing as a PDF – a comparable result instead of loose notes.
The recruiter takes over
The recruiter view holds requirements, questions, scoring and a live transcript. For interviews a human runs, there is a separate stage.
How the agent is built
Each layer handles exactly one job – the audio, the understanding and the conversation are cleanly separated.
Audio layer over LiveKit
LiveKit provides the real-time call room the agent joins. Audio flows in and out through it, stable and with low latency.
Speech recognition and output with Deepgram
Deepgram handles speech-to-text and text-to-speech. What is said becomes text for the agent, and its reply becomes speech again.
Conversation handled by Claude
Claude runs the conversation tailored to the role, based on the CV and the interview dimensions passed in – and keeps it on track.
Recruiter interface
Requirements, questions, scoring and a live transcript in one place. Alongside it, a separate stage for interviews a human runs.
Illustrative example – no real candidate data.
From conversation to a comparable result
Every interview ends with scores along the dimensions and a detailed briefing as a PDF. On top of that come pipeline features: reminders for stalled applications, a view of waiting candidates, and analytics across the interviews conducted.
Before vs. after
Before – manual screening
- First interviews tie up recruiter time
- Results as loose notes
- Candidates hard to compare
- Screening does not scale
After – guided & structured
- First interviews run automatically
- Score and PDF briefing per interview
- Comparable results instead of notes
- Cost per interview in the low single digits
Who it pays off for
Recruiting teams with many first interviews
That want to automate a large share of recurring screening calls without giving up a real conversation.
Platforms scaling screening
That run interviews at volume and need comparable, structured results.
Teams that want fair comparability
Score and briefing per interview instead of loose notes – as the basis for the next stage with a human.
Questions & answers
It joins the call room and runs the first screening interview tailored to the role. After the conversation it delivers scores along the interview dimensions and a detailed briefing as a PDF.
A real conversation. The role, interview dimensions and language are passed in when the room is created, the agent loads the CV data and responds to it – instead of working through a fixed list of questions.
The audio layer runs over LiveKit, speech recognition and output over Deepgram, and the conversation itself over Claude. Each layer handles exactly one job.
Yes. The recruiter interface holds requirements, questions, scoring and a live transcript. For interviews a human runs, there is a separate stage.
The cost per interview conducted is in the low single digits.
Yes. On top come pipeline features: reminders for stalled applications, a view of waiting candidates, and analytics across the interviews conducted.
A process that keeps repeating?
Whether it is a voice agent or another workflow – let us look in a free call at what can be automated in your case.