A workhorse for structured ML.

Seabiscuit was originally created to be the engine inside Equimorphic, a platform built for extracting horse back profiles for saddle fitting. The development process that went into the making of saddle fitting algorithms was heavily experiment driven, which meant that the ability to say "Hey let's try it like this!" and get detailed results on what parts worked and what parts did not a few minutes later rather than after days of reworking code was absolutely pivotal. They needed a unified ML toolkit that could visualise 3D data; represent a diverse panoply of algorithms pithily; and automatically give deep insights into every step of execution.

Equimorphic cofounder Gabriel Jordaan took to the challenge with delight and set about building such a platform. Today, Seabiscuit is a fully differentiable (well, mostly differentiable, but we have evolutionary algorithms for when it's not) language where every operation compiles to an inspectable and trainable graph. Equimorphic were the first to use Seabiscuit in production, but its uses go far beyond horses.

It fits any algorithm where you know or suspect the general shape but need parts that can be learned (assembly robotics, anyone?), neural networks you want per-layer diagnostics on without writing the plumbing yourself, and even purely symbolic routines that just need to be visualised step by step. In each case, you write the part you understand, train the part you don't, and see everything in between.

The language isn't open source yet, but it is available through our early access program.