How the AI interview works — Careers

When you apply, you choose how you are interviewed: by voice with Themis, Metis Foundry’s AI interviewer (the first-round default whenever voice is available), in writing with Themis, or with a person. Every choice is fine and carries no penalty.

What Themis is

Themis is an AI system (Anthropic’s Claude model; in a voice interview it speaks with a stock synthetic voice), not a person. It says so at the start of every interview, out loud and in the captions.
Before it asks any question, it asks your permission to transcribe. In a voice interview you answer by pressing “Yes, transcribe”; your microphone is turned on only after that, and nothing you say is transcribed before it. In writing, transcribing means saving your typed answers. If you say no, you can type instead or have a person interview you.
You can switch between speaking and typing, or ask for a person, at any time, with no penalty. A voice interview takes about 15 minutes of speaking time; in writing there is no time limit, and you can pause and come back with your private link. If the voice interview is not available when you apply, the application page says so and offers the written interview and a person.

Voice: only the words are used

Your speech is converted to text as you speak, and Themis’s questions and your words appear on screen as captions.
We keep no audio recording. Only the text of what you said is stored, encrypted, like a written interview.
Themis does no analysis of your voice, tone, accent, pace, emotion, face or personality, and we create no voiceprint or other biometric identifier. We do not use video.

What is asked and what is scored

Every candidate for a role is asked about the same six topics, in the same order: a practical scenario from the role, a time you handled incomplete evidence, the key sources in your field and their limits, when you would escalate a case, how you handle confidential data, and which of your credentials apply (without giving any numbers). The scenarios differ from interview to interview, with made-up names and figures, but each version is written to be of equal difficulty and is assessed against the same rubric; we record which version you were asked so that a person can check fairness. Themis may ask one short clarifying question per answer.
Only if you give separate permission, Themis scores your answers from 0 to 4 on a written rubric: domain knowledge (25%), the practical scenario (25%), evidence and documentation (20%), escalation judgement (15%), confidentiality practice (10%) and relevant credentials (5%). Each score comes with its reasons and the words it relied on.
With the same permission, Themis writes a short advisory summary for the hiring team with a recommendation (advance, hold, or talk to a person first), quoting your own words, and suggests what a person could ask you next. It never recommends rejecting anyone.
Your name, pronouns, contact details and location are removed before scoring. Spelling, grammar, writing style, speed, length, voice and accent are not scored, and nothing about age, disability, race, sex, religion, national origin or any other protected characteristic is asked about or used.

A person decides

Paul Krzemuski (CEO) or Daniel Maronde (CTO) reads every interview and makes every decision, usually within 10 business days. Scores and recommendations are advisory. No application is screened out, rejected or advanced automatically.
If you live in New York City, your answers are not scored or summarized automatically; a person reviews them.

Your choices, at any time, with no penalty

Speak or type, and switch between them.
Interview with a person instead, before or during the AI interview.
Ask for an accommodation, for example a live conversation, a different format or more support.
Turn automated scoring and the recommendation off.
Ask for a human review of the interview or of a decision, correct something you said, or ask for a copy of your answers and any scores and recommendation and how they were produced.
Ask us to delete your application: we lock it away at once and erase it when the legal four-year retention period ends.

Use your private link (you get it when you apply), or email paul@rocketship.co.

Your data

Answers are encrypted before storage and only the two decision-makers can open them. They are processed by Anthropic, through OpenRouter, only to run the interview, scoring and summary, and are not used to train AI models. In a voice interview, LiveKit and its speech providers convert speech to text and text to speech. We keep no audio, and we have turned off session recording and the speech-to-text provider’s model-improvement program; the providers’ own retention terms apply. We keep application records for four years, then delete them. Full details: the applicant privacy notice.