AI Agent Chains Two Hugging Face Spaces to Build a 3D Paris Monument Gallery With No Manual Asset Creation

Hugging Face engineer Mishig Davaadorj tasked a coding agent with building an interactive website showcasing Paris monuments as 3D Gaussian splats — without touching an image generator or 3D reconstruction tool himself. The agent accomplished this by reading each Gradio Space's auto-generated agents.md file, which provides a plain-text schema for calling and polling any Space via HTTP. It chained two Spaces sequentially: one to generate dark-background specimen images of each monument, and VAST-AI's TripoSplat to reconstruct a .ply Gaussian splat from each single image.

Beyond asset generation, the agent handled downstream production work: detecting and correcting TripoSplat's Y-down coordinate convention, auto-framing each monument, compressing .ply files to .ksplat for faster loading, and building a Three.js viewer with scroll-to-switch and drag-to-rotate controls before deploying the result as a static Space. Davaadorj's only inputs were aesthetic — adjusting zoom, swapping monuments that reconstructed poorly, and tuning transition timing. He frames the result as an early example of a "building-block economy" for multimedia AI, where agents assemble proven, callable model Spaces the way they already assemble code libraries.