# Inkforge — Cline AI Rules

## Project

Inkforge: LSTM+MDN stroke-level handwriting synthesis engine. NOT a font renderer.

## Stack

- Backend: FastAPI 0.110+ / PyTorch 2.x / Celery + Redis / CairoSVG + Pillow

- Frontend: React 18 / Vite 5 / Zustand / Tailwind CSS 3

- Language: Python 3.10+ (backend) / JavaScript JSX (frontend)

## Critical Context

- Strokes are (Δx, Δy, p1, p2, p3) tuples — relative deltas, NOT absolute coordinates

- p1/p2/p3 are mutually exclusive pen states (down/up/end-of-sequence)

- Model: 3× LSTM (hidden=512, dropout=0.2) + MDN (M=20 bivariate Gaussians)

- Style conditioning: latent vector z ∈ ℝ¹²⁸ concatenated at each timestep

## Code Style

- Python: type hints on ALL functions, Google docstrings, PEP 8, 100-char max

- React: functional components + hooks, no class components, Zustand for state

- Tailwind CSS for all styling

## File Organization

- `backend/app/ml/` → PyTorch model code

- `backend/app/models/` → Pydantic request/response schemas

- `backend/app/api/routes/` → FastAPI route handlers (always async)

- `backend/app/services/` → Business logic (inference, rendering)

- `frontend/src/components/` → React components (PascalCase.jsx)

## Rules

1. Never use synchronous inference in API routes — always queue via Celery

2. Never commit .env, checkpoints (.pt), or data files

3. Use config YAML for all hyperparameters — no hardcoded values

4. Conventional Commits for all commit messages