Hiring debrief room where every rating shows its evidence

I want an interview debrief hub, so that the hiring panel walks into the decision meeting with evidence instead of memories. Build it as an app on Greenhouse, Google Meet and Google Calendar. Two kinds of people use it: interviewers who sat on a loop, and the facilitator who runs the debrief meeting. The whole premise is that a competency rating should never be just a number, it should sit next to the thing the candidate actually said that produced it.

The main view lists candidates who have finished an interview loop. Build it from Greenhouse: List Applications to get active applications, List Candidates and Get Candidate to resolve names and the role they are up for, and List Scheduled Interviews to get the interviews in each loop with their times and interviewers. Treat a loop as finished when every scheduled interview in it has already happened. Each row shows the candidate, the role, the stage, how many interviews were in the loop, how many scorecards are in so far, how many transcripts have been captured, and whether the loop is ready to debrief. Sort so the loops that are complete and waiting on a decision come first, and let me filter by role and by whether feedback is still outstanding.

Opening a candidate shows every interview in the loop side by side. Each interview panel shows the interviewer who ran it, their submitted scorecard from Greenhouse List Scorecards with the rating on each competency and their written comments, and the transcript of that actual interview. This is the core screen, so make the side by side comparison comfortable to read across, not a set of collapsed accordions.

Matching a scheduled interview to the right transcript is the part that needs care, so do it in this order. Take the interview start and end time and the interviewers from Greenhouse List Scheduled Interviews. Use Google Calendar List Events over that window to find the calendar event for that interview, matching on the time and the attendees, and read the conferencing details off the event to get the meeting code or link for the call. Then use Google Meet List Conference Records, filtered by that space or by the time range, to find the conference record for the call. From the conference record use List Transcripts to get the transcript, and List Transcript Entries to get the actual speech segments with their speaker and timestamps. Use List Participants on the conference record to work out which speaker is the candidate and which is the interviewer, cross checking participant identities against the interviewer on the Greenhouse interview. Matching is never perfect in practice, so always show which call was matched to which interview, and let someone relink an interview to a different conference record by hand when the guess is wrong and remember that correction.

Inside each interview, present the scorecard as competency rows. Each row shows the competency, the score the interviewer gave it, their written justification, and next to it the candidate's own words on that topic pulled from the transcript entries for that call. Let me click any competency row to jump to that moment in the full transcript, and let me read the full transcript on its own when I want the whole conversation rather than the extracted passages.

Add a debrief mode for the facilitator. It stacks all the interviewers' ratings in one grid: competencies down the side, interviewers across the top, every score in the cells, with the overall recommendation from each scorecard in a summary row. Make disagreement visually obvious, so a competency where one interviewer said strong hire and another said no hire is the first thing the eye lands on. The facilitator opens this before the meeting starts, so the agenda becomes the places the panel actually disagrees rather than a round robin of everyone repeating their notes.

Build in an anchoring guard, and treat it as a hard rule rather than a preference. An interviewer must not be shown other people's transcripts or scorecards until their own scorecard for that candidate is submitted, because reading them first is exactly the anchoring that structured hiring is meant to prevent. Identify the signed in user, map them to their Greenhouse user through List Users by email address, and check Greenhouse List Scorecards for whether that person has submitted their own scorecard on this application. Until it is in, they see only their own interview and their own transcript, and the rest of the loop is visibly locked with an explanation of why rather than silently hidden, so the rule teaches the practice instead of looking like a bug. Facilitators and recruiters are exempt and can always see the whole grid. Once someone's scorecard lands, unlock the full loop for them.

Add a build the debrief packet button that kicks off a background agent for that candidate. The agent reads every transcript and every scorecard for the loop, and for each competency pulls the supporting quotes from what the candidate actually said, attributing each quote to the interview it came from. Its most important job is the honest part: it explicitly flags every place where a scorecard rating is not backed by anything the candidate actually said, whether that is because the topic never came up in that interview, because the transcript is missing, or because the rating simply is not supported by the conversation. Tell it not to soften or invent evidence to fill a gap, since an unsupported rating is the single most useful thing a debrief can surface. The agent also calls out the competencies where the panel disagrees and shows what each interviewer heard. It writes the finished packet back into the app so the candidate page shows it, and posts it to the candidate's profile with Greenhouse Add Candidate Note so the record lives in the ATS. Greenhouse attributes writes to a real user, so send the On-Behalf-Of header with a Greenhouse user id from List Users. Show the run progressing in the app, since reading every transcript in a loop takes a while.

One constraint that shapes the storage design: Google Meet deletes conference records 30 days after the call ends. The app must capture interview transcripts into its own store promptly rather than assuming old loops are still readable. Whenever the app sees a conference record for an interview, pull the transcript entries and persist them immediately, and run a catch up pass over recent loops so transcripts get captured well inside the window rather than at debrief time, which is often more than 30 days after an early screen. Show each interview's capture state honestly: captured and stored, not yet captured but still within the window, or no longer available because the window closed before the app ever saw it. Never render a loop whose transcripts aged out as simply empty, because that reads as a broken app rather than a retention limit.

Persist the captured transcripts, the interview to conference record matches and any manual relinks, the generated debrief packets, the competency sets per role, and the facilitator assignments in the app's own storage, keyed by the Greenhouse application and interview ids. Ratings and scorecards themselves stay owned by Greenhouse and are read fresh, so the app never disagrees with the ATS about what someone submitted.