Neural is a local-first, WebGPU-accelerated neural graph database. It allows you to build, train, and visualize neural networks that live entirely on your device, with no cloud dependencies.
- WebGPU Backpropagation: Custom WGSL kernels for massively parallel forward/backward passes.
- WebNN (NPU) Inference: Direct access to Neural Processing Units for battery-efficient, native-speed execution.
- Real-Time Training: Interactive stochastic gradient descent (SGD) engine running at 60fps.
- Batch Processing: Supports mini-batch training with 3D compute dispatch.
- 3D Interactive Connectome: Render 10k+ neurons and synapses using Three.js and InstancedMesh.
- Force-Directed Layout: Physics-based arrangement of the neural graph in real-time.
- Live Lesioning: Click on synapses to cut them and see the network recompile instantly.
- Local-First: Built on Dash (SQLite + OPFS), storing your neural graph privately on your device.
- Semantic Tagging: Neurons support vector embeddings for semantic search.
- Model Zoo: Explore preset architectures (XOR, Recurrent Loops).
- Import/Export: Share your neural graphs via JSON.
- Brain-Computer Interface: Inject real-time audio (Microphone) directly into the neural network's input neurons.
This project is a high-performance Bun Monorepo:
packages/engine: The Core.GPUEngine: manages WebGPU buffers & pipelines.Translator: Flattens graph topology into dense matrices.NeuronRepository: Persistence layer.
apps/web: The Interface.- Next.js 16 App Router.
- React Three Fiber (R3F) for visualization.
- Glassmorphism Design System.
- Bun v1.0+
- A WebGPU-enabled browser (Chrome 113+, Edge, Firefox Nightly).
# Install dependencies
bun installcd apps/web
bun devOpen http://localhost:3000 to see the Transparent Brain.
Measure your GPU's inference and training throughput:
cd packages/engine
bun run bench- Runtime: Bun
- Frontend: Next.js, TailwindCSS, React Three Fiber
- Compute: WebGPU (WGSL) + WebNN (NPU)
- Storage: @buley/dash (WASM SQLite)
MIT