Pendulum

We started working with Page Nineteen, an applied multimodal AI lab back in 2025, just after they’d graduated from a South Park Commons residency. Our partnership covered a lot of ground, prototyping a generative image canvas, node-based model training environments, and a document intelligence agent together in less than a year. But there was a common thread to each dive, and that was to solve for how creative professionals could actually employ AI into their practice, from model orchestration to building intuitive and intelligent interfaces.

At its core, Pendulum is an intelligent system for indexing, organising and retrieving multimodal assets. We worked with international agencies like R/GA, WPP, and Kering as early design partners, so from day one the aim was to ingest and process assets at large enterprise scale. After ingestion, what did a great user experience amongst thousands of assets actually look like, and what kind of relations between things could be surfaced? We considered all of this across the stack, from engineering sub-second query times on the database layer to an interface that encouraged visual ‘flow’, jumping from one asset to another. We worked closely with both Zo (CTO) and Will (CEO) as one team on pendulum, turning it from idea to prototype in just a couple of weeks.

Every person working in the creative industries knows the importance of a mood board. Mood boards aren’t there just to lift copy from, they are a way of bringing together concepts from different modalities and building a world for a brand to inhabit. With Pendulum, we wanted to build something that could accelerate the process of defining and refining a mood, augmenting as opposed to replacing the chaotic and non-linear process of design and branding.

Pendulum explores a collaborative search UX by balancing what the model views as similarity, which might be something like composition, or by generated tags, which allow a much more surgical exploration of an image library. A user might start with a colour from a brand deck, find a concept they like that fits, and then filter for images that explicitly contain that concept plus images that are hand-drawn, or lino printed. Once a user finds something that resonates, then they can go broad again, feed it in to find a handful of similar references based upon CLIP similarity.

We built each layer of Pendulum from the ground up - a machine-learning ingestion pipeline, the search API and database, and then a design system and interface that ties them into one tool. Projects like this are where we really like to live, splitting our time between implementing technical search infrastructure on the back-end and design craft on the surface.

No single signal is enough on its own - text understanding misses colour, colour misses meaning - so the real work was fusing them into results that actually felt right. We did that fusion inside the database itself, blending semantic, keyword and colour rankings in a single query and tuning the vector index until results that once took ten to thirty seconds came back in around a sixth of one, at full scale.

Populating the image library is a distributed pipeline that embeds, tags, measures and colour-analyses every image in parallel and resumes cleanly when a part fails - enough throughput to process 100,000 images in under five hours.

Pendulum shipped to production at over 300,000 images with sub-second search across every mode. Ultimately, Page Nineteen grew in another direction, but building Pendulum was a chance for us to experience the magic of searching across modalities and glimpse a future where asset collections become a medium for an agency to communicate visually with clients and with themselves.