# AMEVA-Forge Full AI Specification & API Reference
> Comprehensive Reference for AI Agents, LLMs, and RAG Pipelines

## Architecture & Principles
1. 100% Client-Side WebGPU: Zero cloud server costs, 100% data privacy.
2. WGSL Compute Shaders: Parallel matrix multiplication, elementwise ops, and gradient backward shaders.
3. PyTorch Drop-in API: Familiar `.backward()`, `requires_grad=True`, `zero_grad()`, `step()`.
4. Zero VRAM Memory Leak: WebGPU ring buffer pooling and weakref lifetime management.

## API Reference
- `forge.tensor(data, requires_grad=False, device='webgpu')`
- `tensor.backward()`
- `forge.nn.Linear(in_features, out_features)`
- `forge.optim.Adam(params, lr=0.001)`
- `forge.optim.SGD(params, lr=0.01)`

## Package & Source
- PyPI Package: `ameva` (https://pypi.org/project/ameva/)
- License: Apache-2.0
- Author: Eunho Kim (@uno-km), AMEVA Open-Source Foundation