AIInterviewer
Log inSign up

Voice interviews, scored with their evidence

Every score comes with the words that earned it.

Candidates hold a real conversation with an AI interviewer. Recruiters get a report where each number is attached to a verified quote and the moment in the recording it came from — not a black-box rating.

Sign up freeLog in
  • No credit card
  • Your own tenant
  • Self-hostable

Candidate

live

Recruiter

Rubric coverage 83%

0.0overall
“I moved us off a single Postgres box and onto partitioned tables before the migration.”12:04

Our vision

Hiring decisions should be defensible, not just fast

Most screening tools speed up a decision without ever showing their working. A candidate gets rejected by a number nobody can explain, and a recruiter has to trust it anyway. We think the number is worthless unless you can see what produced it.

Quotes, not vibes

Every score cites the transcript. Anything the model cannot back with a verified quote is dropped rather than guessed at.

The candidate never sees a score

They get written feedback on how the conversation went. No band, no verdict — enforced in the schema, not by a checkbox.

Measured is not scored

Pauses, pitch and filler words track nerves, not ability. We report them and refuse to multiply them into the result.

How it works

Four steps, and only one of them is yours

  1. 01

    Describe the role

    Add a job and its description. Versions are kept, so you can see what a candidate was actually interviewed against.

  2. 02

    Review the plan

    Questions and a weighted rubric are generated from the role and the candidate's CV. Edit or approve before anyone joins.

  3. 03

    They talk, it listens

    The candidate joins a voice call from a link. No scheduling, no panel, no calendar tetris.

  4. 04

    Read the evidence

    A scored report lands with quotes, timestamps and rubric coverage — plus separate written feedback for the candidate.

Features

What you get

Real-time voice

Speech recognition, the language model and speech synthesis run on a live turn budget, so the conversation flows instead of stuttering.

Grounded question plans

Questions are written from the job description and the candidate's own CV, then checked for grounding before anyone is asked them.

Verified transcripts

The recording is re-transcribed after the call, so every quote a score depends on is checked against what was really said.

Proctoring you configure

Tab switches, pasting and webcam frames, with thresholds per role. Off by default where a false positive would be unfair.

Reports for both sides

A recruiter PDF with the full evidence trail, and a separate candidate PDF with feedback and no score anywhere in it.

Isolation in the database

Every tenant is separated by row-level security in Postgres, so a forgotten filter in application code still cannot leak a row.

The part most tools skip

A score you can argue with

An unsupported number still reads as authoritative, and that is what makes a rejection unappealable. So the scorer verifies each quote against the transcript, drops paraphrases, and discards any score left with no surviving evidence.

  • Weights renormalise over what was actually asked, so a short interview is not silently marked down.
  • Below half the rubric, no overall number is published at all.
  • “Not assessed” is kept distinct from “no hire” — a failed microphone is not a judgement.

System design

4.2 / 5

Described a concrete migration they led, including the failure mode that forced it and what they would do differently.

We were writing 40k rows a second into one table.08:12
Partitioning by tenant meant the vacuum finally kept up.12:04

Illustrative example — not a real candidate.

Run your first interview today

Create an organisation, add a role, and send a candidate a link. Nothing to install on their side.

Sign up freeLog in