Engineer builds a true-scale universe atlas with AI in one week

A software engineer used Claude Code paired with Fable to build a browser-based atlas that renders the entire observable universe at real measured scale, distance, and speed, from the cosmic web down to quarks. The project incorporates 8.4 million real Gaia DR3 stars with proper motions, 2.6 million SDSS galaxies at true distances, real planetary orbits, and live satellite tracking, all running on a 90 kB gzipped engine written in raw WebGPU with zero runtime dependencies and no game engine.

Across roughly a week, 92 merged pull requests and 237 commits produced about 14,500 lines of TypeScript and WGSL. The AI wrote essentially all of it: the renderer, orbital mechanics (Kepler solvers, SGP4 satellite propagation), a ray-marched atmosphere, and gravitational lensing around Sagittarius A*. The human's role was reviewing PRs, flying through the scene to hunt bugs, and making architectural calls.

What enabled the pace was a verification-first workflow rather than blind trust: planet positions are checked against JPL Horizons in CI and fail past 0.2 degrees of error, data generators refuse to write physically invalid tiles, and CI renders the actual WebGPU scene on software Vulkan for pixel-diff comparison against baselines. Since every camera view is a deterministic URL, visual bugs could be reproduced headlessly and bisected just by pasting a link. The source is open under MIT.