The desktop app
Two ways to run an interview.
Most candidates will never install anything — they open a link and type. The desktop app exists for the sessions where it matters who is actually answering.
Browser portal
Nothing to install. Open a link, answer questions.
This site, at /interview
- Candidate opens an emailed interview code — no account, no download.
- Answers are typed. Questions adapt from the answer just given.
- Camera and microphone are checked before starting, so a dead mic is found up front rather than three questions in.
- Identity is not verified. There is no enrolled reference photo to compare against, so the portal does not claim a check it cannot perform.
Desktop app
A spoken interview, with identity re-checked throughout.
Installed on the interviewing machine
- A reference face is enrolled before the session. A session cannot start without one.
- Identity is re-checked on a jittered interval for the whole session, not once at login.
- The interview is spoken. The model listens continuously, and the candidate can interrupt it mid-sentence.
- The language model runs locally. No API key, and the session works with no internet connection.
Why verification is a separate surface
Conventional proctoring checks identity once, at sign-in, and then watches the room. After that first check, a substitute sitting the interview is invisible. The desktop app re-checks who is producing the work for the duration of the session.
Checks that could not be performed — no face in frame, a closed camera, an unreachable service — are recorded as unmeasured. They are never recorded as passes, and the report says how much of the session was actually covered.
What both surfaces share
The same claim pipeline runs underneath either one. Claims are extracted from the résumé and must appear verbatim in it to survive. Questions are planned against those claims. Evidence is quoted from the candidate's own answer. Neither surface produces a score, and neither makes a hiring decision.
Sessions from both land in the same place — Reports — as one row per claim, with the question that was asked and the answer that did or did not substantiate it.