Recast
The model was already trained. Authorship had to be built
One style held across more films than anyone had time to watch
Someone opens a delivery app at midnight, fills a cart, and closes the tab.
No complaint, no drama, just a quiet exit repeated across a whole country every single night. The standard fix was already sitting on the shelf: a push notification, a discount code, a retargeting ad. The harder brief sat underneath it. Make that one quiet exit feel like it was noticed, for every abandoned cart, on the night it happened. And build the thing to keep pointing at abandoned carts long after the launch that funded it, adaptable to whatever the business threw at it next.



Personal ruled out a person.
At any volume, a human sign-off was the one step that couldn't scale, and scale was the actual brief, not the traffic from a single launch week. So the pipeline had to run itself: a brand-voice language model writing the night from whatever sat in the cart, an image model holding one illustrated world across every render, synthetic narration reading it aloud, a renderer stitching the pieces into a film. Four handoffs, and not one of them had a person standing over it. A tone slip, a broken pose, a frame that drifted off-brand - any of it could ship to a phone at 2am on day one or on the thousandth day the system ran, and no one would be there to catch it first.

Give the pipeline the one thing a person couldn't be at that volume: a reviewer that never got tired.
The language model worked from a fixed brand-voice brief, so the narrative held one register no matter what landed in the cart. The image model rendered against a locked style anchor instead of reinventing itself each prompt, so the illustrated world kept obeying its own rules, render after render. And every frame that came out of it had to clear an automated consistency check before anything downstream touched it. Fail, and it went back to be regenerated and re-scored, quietly, before a person ever knew there'd been a miss.



52,000+ personalised films in the launch's first two weeks alone, each pulled from a different cart, none of them previewed by a person before delivery.
16.5% cart recovery, against an industry norm that discount-driven retargeting rarely clears 5%. The two weeks were the opening burst, not the ceiling: the same gate, the same locked style anchor, and the same standing pipeline kept running 24/7 well past the launch window, ready for whatever triggered it next.





