Show HN: Tesseract Core – Autodiff-native, self-documenting software components

15 points

a year ago

Tesseract Core is a command line app + Python SDK that lets scientists and engineers wrap complex scientific and machine learning code into containerized, self-documenting, differentiable functions—making them easier to use, compose, serve, share, and deploy.

Its main use case is to build end-to-end differentiable pipelines consisting of wildly different components like physical simulators, geometric operators like differentiable meshers / renderers, generic data transforms, and neural networks. Just installing these heterogeneous components can be a challenge for scientists who simply want to get things done—let alone integrating them into large-scale, optimization-driven workflows.

We’re deeply inspired by the way JAX primitives work to expose their functionality [1]. Tesseracts are taking this one step further by building on top of a modern Python stack with Pydantic at its core, used as a DSL to define schemas for expected inputs and outputs (+ some custom sugar to handle arrays and differentiability) [2]. One user-given schema informs all other endpoint schemas, i.e., simply by defining the schema for apply we also generate one for jacobian, vector_jacobian_product, abstract_eval, ..., via some Pydantic metaprogramming [3].

This allows R&D staff to provide minimal config and essentially auto-generate a Docker container wrapping their software in a way that’s ready for RPC (locally and over the network).

[1] https://docs.jax.dev/en/latest/jax-primitives.html

[2] https://github.com/pasteurlabs/tesseract-core/blob/675e07c95...

[3] https://github.com/pasteurlabs/tesseract-core/blob/675e07c95...