Shaderwrap
Full-stack AI tool that translates plain English into production-ready HLSL shader code, powered by a custom fine-tuned 14B parameter LLM trained on hundreds of curated URP shaders.
Overview
Writing HLSL shaders from scratch is slow and sparsely supported by AI tooling. ShaderWrap bridges that gap — describe the visual effect you want in plain English, get production-ready Unity URP shader code back. Built around a custom fine-tune of Qwen2.5-Coder (14B parameters, 5-bit quantized to ~10GB) trained on a dataset of hundreds of open-source and original URP shaders I curated — giving the model domain-specific accuracy the base model lacks. Architecture: FastAPI backend, JavaScript/HTML frontend, local inference via Ollama with cloud API fallback. Solves the real workflow problem that shader authoring is slow, underdocumented, and poorly served by general-purpose AI models.
Gallery
Technical Specs
- Timeline
- November 2025
- Stack
- Python FastAPI JavaScript HTML CSS