This is the first in a series of posts introducing Toymodel. What started as a project to build a financial optimization and prediction engine morphed into a new kind of scientific computing platform. It features its own auto-differentiation engine and composable, modular graphs with nested differentiation, designed to build, inspect, reuse, and share complex ML and optimization models (and more).

The public beta is out now.

More to follow.

[And yes, it solves Rockafellar–Uryasev CVaR LPs in its sleep, but that’s beside the point now.]