Companion · Concept D
Concept D · interactive · a feed of questions

The companion that reads along

A short brief sits at the top and expands only when wanted. Below it, three questions to start. The companion keeps listening, but it only writes new cards when the interviewer asks for them: tap a chip, Technical, Experience, Culture or Wrap-up, and it writes the request into the box for the interviewer to adjust and send. Three questions arrive grounded in that source, with a small Generate 3 more link under them. The transcript is behind a toggle, not in the way.

Chrome · Google Meet + Evalzo side panel
evalzo Companion
Listening
Maya Patel Full Stack Engineer · First round
Transcript · final utterances only
Listening
Ticked questions and requests are recorded as interviewer activity. They do not change candidate scores.
Review controls · not part of the product. The transcript replays automatically; cards appear only when you generate.
What changed from concept A

Three at a time, on request

Concept A kept a single cue and a Next button. Here the panel opens with three questions and generates three more only when asked, so inference runs once per batch rather than on every utterance. Four prompt chips fill the box rather than firing: Technical from the requisition, resume and recent answers; Experience from the resume and the prior round; Culture from the criteria; Wrap-up from the plan. The interviewer can edit the request before sending. Generate 3 more sits under the batch it extends, inside the list. The listening line counts answers heard since the last technical batch.

The brief collapses to one line with a flag count. Open it when a reminder is needed; it never pushes the questions down for long.

Steering, with a relevance check

Requests go through the same grounding as cards

  • In requisition and resume (Java, PostgreSQL): cards immediately, with sources.
  • In one of the two (AWS, React): cards with a caveat about what the other source says.
  • In neither (C++): the companion says so and asks before generating. Those cards are dashed and recorded as off-plan when ticked.
Limits to test

Not yet validated

  • Flags come from the written assessment record and resume-to-requisition comparison. Their wording needs a technical reviewer's sign-off so they read as prompts, not verdicts.
  • The transcript is behind a toggle. If pilots never open it, remove it.
  • Batch latency is simulated. A batch should return in under a few seconds, or the interviewer will stop asking for one mid-conversation.