Tinted

Lighting is the dominant error source, and correcting it deterministically beats hoping a model learns invariance. LAB separates lightness from color so classification works on the right axes.

Monk was built for inclusive tech, with real coverage of deeper skin tones. That's the exact failure mode this project exists to avoid.

CLIP matches visual similarity without labelling thousands of products, and precomputing the whole catalogue reduces per-request work to one matrix multiply. There's a Lab-distance fallback if the model fails to load, so the feature degrades instead of dying.

Blur, over/under-exposure, yaw and pitch are each checked with a specific, actionable error message. For a product whose whole purpose is getting tone right, a clear refusal beats a confidently wrong reading. The pitch estimator is still marked provisional in the code pending calibration across the eval set.

17 landmark patches are sampled on flat malar and forehead skin, then discarded if they read as an edge, a specular highlight, or deep shadow, with a 10 percent trimmed mean over what remains. If fewer than five patches survive, the request fails rather than guessing.

Per-minute and per-day limits with correct Retry-After headers and bounded memory, on top of a cap that meters real token usage against model pricing and stops at a dollar ceiling. A public demo calling a paid API needs a hard spending floor, not good intentions.