NeuroBranch
A visual, node-based editor for composing PyTorch neural network architectures as typed, executable cards, with two-way sync to generated code, step-by-step tensor execution, and a constrained AI planner that proposes only valid graph changes.
NeuroBranch replaces a wall of architecture-defining Python with a graph of small, typed, executable cards on a canvas. Dragging and wiring compatible cards generates readable PyTorch code, and edits to supported PyTorch code flow back into the graph, keeping both representations in sync. The graph can be run, rerun, reset, or stepped through tensor by tensor in a local Python runtime.
An "Ask NeuroBranch" copilot plans an architecture described in plain English, wiring only compatible ports and handing back an auditable plan to review or auto-apply, using the same validated tools and ports available to a person. Custom, reusable nn.Module cards can be built in a dedicated studio, and drafts, custom cards, optimizers, and presets persist locally in SQLite.
It is an early, evolving personal project built as an Electron desktop application with a React and TypeScript frontend.