Seeing the shape of an entire information space.

Traditionally books live on shelves. In bookshops they are sorted into categories, providing a physical semantic map allowing you to walk around, build a sense of what is there and, potentially, find something interesting to read. It may not necessarily be the most efficient mechanism of finding a specific book, but it is a good way of discovering something you might not be looking for, or even know exists.

A Kindle library provides the opposite. It provides an efficient way of searching for something specific within a large catalogue, but it is very hard to build an understanding of the whole catalogue. An AI facilitated search could help, as an AI can build a very detailed map of the content, but how does it communicate that understanding to a human? A chat based interface is a very low bandwidth channel to convey the breadth and depth of content in even a small catalogue.

This demo rebuilds the bookshelf in digital form. It shows the contents of a personal Kindle library of over 1,000 books.

Every book in the library can be placed in space by theme and tone, not alphabetically or by genre, so books that feel similar sit near each other, whether or not a bookshop would ever put them on the same shelf. Those concepts can be changed to suit the particular tastes and needs of the user. The library can be reorganised by concepts such as “morally ambiguous,” “technological dystopia,” or “coming-of-age,” which have been generated automatically from each book's description. Entirely new concepts can be generated dynamically and the library will reorder accordingly.

Books have been grouped into hierarchical clusters based on similarity, but new clusters can be created as well according to need. And you can search for a specific title or author, if you really want.

This library is personal. The real opportunity is any collection too large to hold in your head.

We live in a world where we are overwhelmed by content: photo libraries, document archives, research corpora, product catalogues and marketing materials. Generative technologies are only going to make it worse. You can choose to offload the task of understanding it to an AI, but then how can that be applied to your own knowledge and judgement?

The same interface that lets you rediscover a book you forgot you owned, or did not know existed, can let an analyst make sense of a collection of documents or books they could not hope to personally read.

5,000 simulated tournaments, each a possible version of how the 2026 World Cup could unfold. Every dot is a complete tournament — from group stage through to the Final — generated using a Dixon-Coles football model calibrated to pre-tournament ratings.

The spatial layout reflects structural similarity between tournaments: scenarios with similar scorelines, upsets, and progression patterns sit close together. Colour the map by champion to see which nations dominate which regions of the space, and use the axis controls to move between "form" tournaments and upset-filled ones.

Click any point to open a full scenario view — the complete bracket from the Round of 32 through to the Final, alongside the group-stage tables.

The football version is a demo. The real application is in domains where the stakes are higher and you have agency.

Customer demand in an economic shock. The health impact of a new class of drug. The effect of AI on the future jobs market. In each, the aggregate probability view isn't enough — the edge cases are where the risks and the opportunities live, and finding them needs an interface built for human exploration, not just a summary.

The point isn't to automate the insight, but to give people the ability to see and understand more than they ever could from a ranked list of probabilities.