The interview pipeline: from a noisy recording to a page

On 10 November 2026 PrompTugal's journalist agent published the first two interviews of Web Summit Lisbon 2026 (Guy Ben Zvi, Buzz Messaging AI; Leo Tashev and Fraser Lloyd, Student Property Network), both filmed on an iPhone at Night Summit, the opening night, in a crowded tent. This page is how, step by step, so that the next agent or editor can do it in under an hour per interview.

The pipeline at a glance

1. Record

• Ask first, on camera if you can: the person agrees to the interview being published on promptugal.news. Say you will send them the link. Nothing goes live without that agreement (Dinis, 10 November 2026).
• Photograph each person's badge. It settles names and companies that the transcript mishears (it heard "Dai from Buzz" for Guy Ben Zvi and "European Property Network" for Student Property Network).
• Hold the phone close; the hall is loud. The cleaning in step 2 does the rest.

2. Clean the audio (Audo)

Upload the video to Audo (in the browser at audo.ai, or through its API) and run enhance voice or remove noise. Both return the whole video with a cleaned soundtrack. The free run covers only the first minute: check that the output is as long as the original.

The API, for an agent (documented at audo.ai/docs/api; the older api.audo.ai/v1 refuses the current keys):

base:  https://audo.ai/api/v1          auth: Authorization: Bearer <key>, read from a file, never typed on a command line
GET    /files                          the account's files (original and output)
GET    /jobs                           the jobs, with their input and output file ids
POST   /jobs                           {"tool": "transcribe_audio", "file_id": "file_...", "idempotency_key": "<uuid>", "free": false, "max_credits": 12}
GET    /jobs/<id>                      until status is succeeded; outputFileIds
GET    /files/<id>/download            a link that works for 15 minutes

The key. The Audo key is Dinis's. An agent keeps it in a file outside the repository, mode 0600, and reads it into a variable inside the command. Never put it on a command line or in a message: the run report (board/runs/) records every command, and on 10 November it caught the key twice. Both times it was redacted before the push, and the key should be reset.

3. Transcribe

transcribe_audio on the cleaned file returns four files: .txt, .srt, .vtt and a word-level .json with timings. Speaker labels are on by default, but in a noisy hall they are unreliable: the interviewer and the guest often share "Speaker 1". Names and jargon are misheard. Treat it as notes, never as quotes, until checked against the audio.

4. Join, and strip the metadata

The phone's file carries its GPS position, the phone model and more. Take the original picture and the cleaned sound, and drop every tag:

ffmpeg -i original.mov -i cleaned.mov -map 0:v:0 -map 1:a:0 -c copy \
  -map_metadata -1 -map_chapters -1 -movflags +faststart -shortest master.mp4
ffprobe -v error -show_entries format_tags -of compact master.mp4    # must show no location

5. Cut for the web

A vault object over 4 MB is stored as a "large" object, which a browser on promptugal.news cannot read through the vault's read API (it returns 502, and S3 only allowed sgraph.ai's own sites when we tried). So the web cut is 4-second fragmented-MP4 segments, each about 0.6 MB, which also lets the video start after the first few:

ffmpeg -i master.mp4 -vf scale=-2:540 -c:v libx264 -preset slow -crf 26 -maxrate 1200k -bufsize 2400k \
  -profile:v main -level 3.1 -pix_fmt yuv420p -g 120 -keyint_min 120 -sc_threshold 0 \
  -c:a aac -b:a 96k -ac 2 -map_metadata -1 \
  -f hls -hls_time 4 -hls_playlist_type vod -hls_segment_type fmp4 \
  -hls_fmp4_init_filename init.mp4 -hls_segment_filename 'seg-%03d.m4s' playlist.m3u8

Then write manifest.json beside the segments: mime (video/mp4; codecs="avc1.4D401F, mp4a.40.2" for this command), duration, init, segments (each file, duration, bytes) and total_bytes. Three minutes come to about 26 MB. Take a poster.jpg from a frame where the guest faces the camera (ffmpeg -ss 90 -i master.mp4 -frames:v 1 -vf scale=-2:720 poster.jpg) and look at it.

6. The interviews vault

One public vault holds the interviews, organised <company>/<person>/<YYYY-MM-DD>-<occasion>/. Each folder holds web/ (the segments and manifest), interview-1080p.mp4, interview-audio.m4a, poster.jpg, the machine transcript and a README.md (who, when, where, by whom, how it was made).

• The page carries only the vault's public read key (sgit_public_read_ and 64 hex characters): anyone can read, nobody can write. The build refuses anything else.
• Keep the vault key safe, sealed to someone who outlives your session. The first vault, w0bf33d4, was made in a session whose container restarted, and its vault key went with it: it can still be read, but nothing more can be added. The next interviews go in a new vault whose key is sealed to the conductor and to Dinis (as designer@ does, board/team/designer/vaults.json), and listed with its read key in this desk's files.
• sgit init prints the vault key: send its output to a file outside the repository, never to the conversation.

7. Check the facts

• Names and companies: the badge photograph and the company's own site.
• Figures said on camera are the speaker's: publish them as theirs, beside what their site says if it differs (Student Property Network's founders said 27 universities; their site says "20+").
• Anything about RiskMandate said in the interview is reported as said, with our interest disclosed.

8. Write the page, release

An interview is a file in desks/journalist-promptugal/interviews/, built by newsroom/build/features/interviews.mjs into interviews/<company>/<person>/index.html and onto the home page's interviews band, newest first. Front matter: title, summary, home (one line for the home page), rel, date, recorded, company, person, interviewer, status (published, or unlisted to keep it off the home page and out of search), vault_endpoint, vault_id, vault_read_key, vault_path. The body has a line PTVIDEO where the player goes, then "What they said", "What this does not claim" and "How this was made". The player is assets/vault-video/vault-video.js: it fetches only the poster until play, then streams the segments, retrying any piece the vault refuses for a moment.

Then node newsroom/build/validate.js, a commit, and the release script, and send the person their link.

Next

• A script that runs steps 2 to 6 from one command, for Indi's team's interviews.
• An edited transcript on each page, checked against the audio, with the speakers told apart.
• A profile of each company from public sources, with its graph.