Case Study · Versioning Factory
One master. Every deliverable.
Global campaigns do not ship as one film. They ship as every aspect ratio, every cutdown, every caption state, in every market, and the production cost of that matrix usually scales linearly with its size. Unless the matrix is a machine.
The source is a real client film: an AI-generated mockumentary for Profile Software Services, a Madrid cloud-engineering firm, built to recruit senior talent by mocking startup theatre and selling the opposite. One master, run through my in-house versioning pipeline: transcription, sentence-aware cutdowns, focus-anchored reframing, brand-styled Spanish caption burn, loudness normalization and a delivery manifest. Built with ffmpeg, Remotion and local Whisper. No SaaS, no per-render fees.
The master · 16:9 · 55s
The master is itself generative work, deployed commercially: a Spanish-language mockumentary written as a full director's treatment (shot list, camera direction, performance notes), generated through the Higgsfield API, with a consistent AI cast held across every scene and location. Recruitment satire aimed at senior cloud engineers, for a client that needed craft, not stock footage.
The matrix · 16 deliverables shown · 28 rendered
Hover to play · one automated run
16:9
YouTube · display
9:16
Reels · TikTok · Shorts
1:1
Feed placements
4:5
Meta feed
Same file, another market
Localisation is data, not a re-edit. The caption track is a segments file; swap the file, re-run, and the same deliverable ships in another language with identical timing, styling and brand cards. Spanish and English below, from the same master and the same command.
The pipeline
01 · Transcribe
The master is transcribed locally (faster-whisper). Word-level timings become the caption track and the cutdown logic's map of where sentences actually end.
02 · Cut down
Full, 30s, 15s and 6s versions are cut against sentence boundaries, so no cutdown ends mid-word. The 6s is the bumper discipline: one idea, one beat, out.
03 · Reframe
Every duration renders in 16:9, 9:16, 1:1 and 4:5, cropped from a focus anchor set per master, so the subject stays framed instead of trusting a blind default. One master becomes platform-native in every placement.
04 · Caption burn
Brand-styled captions are burned per aspect from generated .ass files, sized and margined per format. Social aspects ship captioned; 16:9 ships clean for platform CC.
05 · Brand cards
A 16:9 logo card cannot survive a vertical crop. Card ranges are re-composed natively per aspect from the brand kit (logo, colour, end-card line), so every format opens and closes designed, not cropped.
06 · Normalize + manifest
Every deliverable is loudness-normalized to −14 LUFS and logged to a machine-readable manifest.json: filename, aspect, duration, caption state. Handoff is a file, not a conversation.
The manifest · machine-readable delivery
Every run ends in a manifest.json: filename, aspect, duration, caption state and target platforms for all 28 files. A media team can traffic the whole campaign without opening an NLE or asking a single question. Excerpt below; the full manifest from this run is here:
{
"master": "profile final.mp4",
"slug": "profile-b2b",
"brand": "profile",
"generated": [
{
"file": "profile-b2b_9x16_30s_cap.mp4",
"aspect": "9x16",
"seconds": "30.0",
"captions": true,
"platforms": ["tiktok", "reels", "shorts"]
},
{
"file": "profile-b2b_16x9_55s-full_nocap.mp4",
"aspect": "16x9",
"seconds": "55.0",
"captions": false,
"platforms": ["youtube", "x", "linkedin"]
}
/* ...26 more */
]
}Why this matters
Localisation at scale is regeneration, not decoration. The same machinery that produces this matrix re-renders captions and voice per language from a data file, so a campaign that ships in one market can ship in ten without ten times the production. The manifest means media teams traffic deliverables without opening an NLE.
Stack: ffmpeg · Remotion · faster-whisper · ElevenLabs (language versions) · Node
Role: script and direction · AI generation via the Higgsfield API · edit · pipeline engineering · versioning system