Interviews you can defend six months later.
Getdio runs the structured interview, transcribes it, and scores every competency against the candidate's own words — with the quote attached. You review the evidence and make the call.
No card required. Bring one job description and you will have a working interview in about five minutes.
Evidence is the product.
Most AI interviewers give you a number. The number is the easy part — defending it is the job.
of scores cite a verbatim span from the transcript. A score with no quote behind it is rejected by the pipeline, not shown to you with a shrug.
independent models score every competency, and they are required to be different families. Two runs of one model produce correlated errors that agree confidently.
résumés, names or photos reach the scorer. It sees a redacted transcript, so what it rewards is what was said.
Three steps, and none of them is "trust the model".
The interview is automated. The decision is not.
Paste the job description
Getdio turns it into a rubric where every competency traces back to a line in the posting. Anything it inferred rather than read is labelled and switched off until you turn it on.
Send a private link
One link, one candidate, expires in 72 hours. They need a microphone and about twenty minutes. No install, no scheduling, no camera.
Review the evidence
Each competency arrives with the quotes that produced it. Click a score, land on the words. Change it if you disagree — that is the point of a human step.
Every score is one click from the sentence that caused it.
The scorer must cite a verbatim span. The span is verified against the transcript character by character, and a citation that does not match is rejected rather than displayed.
- Click a criterion, the transcript scrolls to the quote and highlights it.
- Abstention is a first-class result — a thin answer reads as “insufficient evidence”, never as a low score.
- The composite is suppressed entirely when too little of the rubric was covered.
Two models score it. One human decides it.
Every competency is scored twice by different model families. Where they disagree, the disagreement is surfaced as the signal it is — and no result reaches a hiring decision without a person putting their name on it.
- Disagreement above threshold flags the criterion for review automatically.
- Overrides require a written reason, which becomes the corpus for improving the rubric.
- Every decision records who made it, when, and on what evidence.
The scorer never learns who the candidate is.
Names and résumés are visible to your hiring team and structurally absent from assessment. What gets rewarded is what was said in the interview.
- Redaction happens in the pipeline, not in a prompt asking nicely.
- Biased or ambiguous phrasing in the job description is flagged before you publish it.
- Integrity signals sit outside the assessment and cannot move a score.
Built to be audited.
Hiring is a decision someone may have to justify to a candidate, a regulator or a court. The system assumes that from the start.
One candidate, one link, 72 hours. Revocable instantly, and a revoked link never says “already taken” to someone who never took it.
A published rubric is content-hashed on disk. “Which rubric was this candidate assessed against” has a real answer, years later.
The interviewer explains recording and asks at the start. A candidate who declines is not assessed, and the session says so.
Transcripts and recordings are stored as files you can read, export and delete without going through us.
Questions worth asking.
Does the AI decide who gets hired?
No, and it cannot. A scorecard carries a required human review step in the schema itself — there is no code path that turns a model score into a hiring outcome. Getdio's job is to produce evidence and a defensible first read; yours is to decide.
What does a candidate actually experience?
They open a link, allow the microphone, and have a spoken conversation of about twenty minutes. No camera, no install, no scheduling. They can ask to repeat, skip a question, or stop at any point, and stopping does not produce a bad score — it produces no score.
How do you stop it rewarding confident nonsense?
Scores must cite verbatim spans, and the spans are verified against the transcript. A model that asserts something the candidate did not say cannot produce a citation that matches, so the score is rejected. The summary is separately consistency-checked against the criterion results, and unsupported claims are stripped.
What happens when the model is not sure?
It abstains, and abstention renders as its own result rather than as a low score. If too little of the rubric was covered, no composite is produced at all — a number renormalised over a minority of the rubric is not a score.
Can we use our own rubric?
Yes. The compiler proposes competencies from your job description, and every one of them is editable, weightable and removable before you publish. Nothing is assessed that you did not switch on.
Your next interview can be one you would show a lawyer.
Paste a job description and see the rubric it produces. Nothing is published until you approve it.