Case study Content & social media
A video factory: raw footage in, finished posts out
The brief
Produce short-form video at volume from real footage, without an editor and without AI-generated imagery.
Editing is the bottleneck in social video. Hours of footage hide a few seconds worth posting, and finding them by hand does not scale past a few posts a week.
- 01 Raw footage
- 02 AI catalogue
- 03 Signals
- 04 Edit specs
- 05 Render
- 06 AI QA
- 07 Review
Fig. — System schematic. Raw footage is catalogued by AI, cut points come from signals, posts are specified as data, rendered, checked and reviewed.
What we built
- 01
AI that watches the footage
A multimodal model logs what happens in every clip, picture and sound. Classical signals, such as scene cuts, audio peaks and subject tracking, then pin down the exact cut points.
- 02
Edits as data
Each post is a structured edit spec, checked by a validator for reuse limits and variety across hook, trim, format and audio, then rendered in code with Remotion.
- 03
Quality control before a person looks
An automated QA pass checks every render, and a review gallery presents the batch for approval. A research loop studies what performs in the niche and feeds the editing playbook.
How we worked
- 01 Research
Studied what performs in the niche and turned it into a playbook.
- 02 Pipeline
Transcoding, AI cataloguing, signal extraction and a merged footage digest.
- 03 Production
Edit specs, validation and batch rendering.
- 04 Review
Automated QA, then a gallery for human approval.
What changed
- Batches of 50 to 100 finished posts from one footage drop.
- Real footage and real audio: no AI voices and no generated images.
- A repeatable process with a quality check at every step.