Team Wexy

Wexy (https://wexy.ai) is an agentic WebGPU authoring studio that turns natural-language ideas into real-time, interactive GPU programs in the browser. It helps developers and visual creators prototype procedural worlds without hand-writing every shader and engine interface. Unlike text-to-video systems that return rendered clips, Wexy generates executable programs that users can explore and adjust locally. Meta's models design the scene and write constrained GPU code; Wexy's verification and repair pipeline bridges the gap between generated code and a working interactive world. Structured Planning & Code Generation: Meta Model API translates visual prompts into structured scene designs, rendering decisions, and modular WGSL shader code that targets defined engine contracts. Contract Validation & Automated Repair: Wexy checks generated candidates against engine interfaces and uses compiler and contract diagnostics to guide multi-turn model repair. The model proposes code; deterministic checks determine whether it meets the supported contracts. Local WebGPU Execution: After validation and explicit user approval, candidates run on the user's GPU through NullGraph (TypeScript/Vite/WebGPU). Creators can move the camera, adjust supported scene parameters, and capture MP4 video locally, without a new model request for every viewpoint. Inspectable Agentic Workflow: Judges can explore the studio or inspect the step-by-step developer flow at https://wexy.ai/studio.html?agent=1, including model-call diagnostics, generated artifacts, and candidate checks. Saved worlds and replay fixtures provide additional inspection paths, clearly separated from live Meta inference.