Handwired

Handwired is a tool that turns regular JavaScript code into a neural network. Rather than training a model, it compiles your existing logic into an exact yet fully differentiable form, so the program can be run and differentiated directly in a browser without a GPU.

The core mechanism rewrites branches and comparisons into differentiable primitives; for example, a conditional becomes a combination of relu terms. This preserves exact behavior while exposing gradients. The site includes a tutorial, GitHub link, and demos such as Mini OS, a small desktop written in JS and transformed with Handwired, and a learned gallery where compiled weights can be unfrozen and trained in real time.

Another demo searches 100,000 JSON documents, with regex matching, JSON parsing, and highlighting written in JS and turned into WebGPU kernels for parallelism. These examples show the approach scaling from small snippets to interactive browser applications.

Handwired is aimed at developers and researchers interested in differentiable programming, program synthesis, and browser-based computation. It is not a conventional ML framework and requires comfort with JavaScript and neural-net concepts.

To get started, read the tutorial, try the demos, and inspect the GitHub repository. Handwired is best treated as an experimental platform for exploring differentiable code rather than a production training system.