Interview debrief desk that gets feedback written the same day

I want an interview debrief desk that our recruiters and hiring managers open right after an interview loop, so feedback gets written the same day instead of a week later from memory. Two kinds of people use it: interviewers who owe a writeup, and recruiting coordinators who chase them.

The landing view is a chase list. Show every interview that happened in the last seven days with no debrief filed yet, sorted oldest first so the most overdue sits at the top. Each row shows the candidate name, the role, the interview stage, the interviewer who owes the writeup, when the interview happened, how many days it has been waiting, and whether a recording has been matched and transcribed yet. Let me filter by interviewer, and default the list to the signed in user's own interviews with a toggle to see everyone's.

Build that list from Greenhouse. Use List Scheduled Interviews over the last seven days to get the interviews and their interviewers, then use List Applications and Get Candidate to resolve the candidate name, the role they are up for, and the stage the interview belongs to.

Match each interview to its recording in a Google Drive folder that I configure once inside the app. Use List Files scoped to that folder with a Drive query that filters to audio files created around the interview date, and Get File Metadata to confirm the created time, file size, and mime type. Match on the interview date first, then on the candidate name appearing in the file name. File naming conventions drift in practice, so always show which file was matched, let me relink an interview to a different file by hand when the guess is wrong, and remember that correction.

Each interview has a Transcribe button. The handler downloads the audio bytes with Download File Content and posts those bytes straight to Deepgram's Transcribe Pre-recorded Audio operation. Do not pass Deepgram a Drive link, because Drive files are not publicly reachable and Deepgram cannot fetch them. Send the raw bytes with an audio content type, authenticate with the header Authorization: Token followed by the API key rather than Bearer, and turn on diarization and smart formatting as query string parameters. Diarization is the part that makes the interviewer versus candidate split possible. Transcription is slow and costs money, so store the finished transcript and reuse it rather than re-transcribing every time someone reopens the interview.

Clicking an interview opens a debrief page split into two panes. The left pane is the transcript, grouped into speaker labeled turns with timestamps. Deepgram returns anonymous speaker numbers, so let the user say once which speaker is the candidate, then relabel every turn accordingly and remember that choice. The right pane is a structured feedback form: a short rating per competency, an evidence field under each competency, and an overall recommendation such as strong hire, hire, no hire, or strong no hire. The competency list should be editable per role and saved. The whole reason the panes sit side by side is evidence, so let the interviewer select any passage in the transcript and attach it to a competency as a quote, carrying the speaker and timestamp across with it. Save the form as a draft while they type, so a half written debrief survives a closed tab.

Submit posts the finished writeup onto the candidate's Greenhouse profile using Add Candidate Note, formatted as readable text with the competency ratings, the evidence quotes, and the overall recommendation. Greenhouse requires writes to be attributed to a real user, so send the On-Behalf-Of header with a Greenhouse user id. Once the note posts, mark the debrief filed, record who filed it and when, and drop it off the chase list.

Track which interviewer owes which writeup. Give coordinators a view grouped by interviewer showing how many debriefs each person owes and how overdue their oldest one is, and let a coordinator mark someone as nudged with the date so the same person is not chased twice in one day. Keep the nudging inside the app rather than sending messages out to other tools.

Persist the debriefs, drafts, transcripts, speaker labels, recording matches, competency sets, and nudge history in the app's own storage, keyed by the Greenhouse interview id. Make it obvious when an interview has no recording matched at all, since a missing recording is a different problem from an unwritten debrief and gets solved a different way.