Page Nineteen is an applied multimodal AI lab in London — a founding team of six out of a South Park Commons residency, working with R/GA, WPP and Kering. Their Python toolkit could already see, tag, train and generate; at stake was whether that engine became a product. We led product design zero to one on both of the lab's surfaces: Pendulum, a multimodal semantic-search platform, and Lightnote, the creative toolkit that followed it.
Pendulum came first. In September 2024 the founders brought us in to take it from a thesis with no interface to a working product: we prototyped early, moving from wireframes to working demos within weeks, and tested them in calls drawn from a personal network. It indexed, organized and retrieved millions of images and documents for downstream generative workflows.
The obstacle was evidence. Users leaned in, but the signal would not carry investors, and to push on was to spend real engineering capital on an unproven bet. The call was to set the direction down before that capital was spent — and that decision was the deliverable, because the real function of zero to one is to learn cheaply.
Lightnote began weeks after the pivot, when the lab needed its own fine-tuning tooling. The brief arrived as a UI problem: put panels over the model calls. The real problem was grammar. The CLI already spoke a complete language of sets, processes and runs, and an interface with a vocabulary of its own would have forked the product in two. An early direction gave each model family a bespoke panel — richer on screen, and three dialects of one language.
We rebuilt around one node form with flags as fields, so models differ only in their parameters. Parity is enforced by the API, not by discipline: a trigger word lands in the same call as its --trigger flag, and nothing drifts. The lab did its own work in the canvas we delivered. Operations later paused at the investors' call, and in 2026 Page Nineteen released the toolkit open source under MIT.