Sheaf

Sheaf is a functional programming language designed for differentiable computation, with a focus on machine learning applications. It introduces a code-as-data paradigm inspired by Clojure, enabling models to be represented as inspectable, composable, and compiled data structures. This approach is intended to support machine learning researchers and those working on agentic AI by streamlining the process of defining and manipulating neural networks.

The language emphasizes mathematical clarity and resource efficiency, allowing users to write mathematical expressions directly without boilerplate code or class structures. Sheaf provides runtime observability features, such as the ability to catch NaN values, trace tensor shapes, and profile performance without requiring modifications to the source code. Neural networks in Sheaf are structured as compositions of mathematical functions over parameter trees, aligning with a purely functional programming model.

Sheaf is delivered as a single binary executable with no external dependencies, and it supports training and running models on GPU hardware out of the box. The language also features a uniform syntax, using a single syntactic form for all operations to minimize ambiguity and reduce generation errors. For users onboarding with AI code generation tools such as Claude Code, Cursor, and Copilot, Sheaf includes a built-in context generator. The language claims to offer significant reductions in code verbosity, with 60-75% fewer tokens than equivalent Python implementations for the same machine learning architectures.

Sheaf targets machine learning researchers and developers seeking a functional, data-centric approach to building and inspecting neural networks, as well as those interested in agentic AI.