Dev uses Claude to label 200 unsupervised visual clusters in a painting-retrieval app
jamievurnilla · reddit · 2026-09-20
A developer built a painting-retrieval site where you mark a point of interest and get eight similar regions from other paintings. A custom vision model encodes 308 paintings into hundreds of thousands of patches clustered into 200 unlabeled groups.
- Claude inspected the first 16 crops from each cluster and described them, naming clusters humans struggle to label (e.g., necklines and collar-shoulder junctions)
- More valuable: Claude flagged 3 clusters tracking picture frames and scan margins instead of painted content, which the author demoted before shipping
- 3,200 crops reviewed in total, impractical to check by eye
More from coding & agent
- Jens Honack: Jev Is Like a Smart Switch Statement, More Dynamic Than Low-Dimensional Determinism — JensHonack · 2026-09-20
- Teknium Critiques Tamara's Jev Compaction Strategy: Valid Use Cases, Wrong Approach — Teknium · 2026-09-20
- Swapping a 13s pipeline step for a 200ms call saves thousands per month — hardimanjames · 2026-09-20
- Opus 5-built Three.js WebGPU waves deliver stunning real-time shoreline in browser — majidmanzarpour · 2026-09-20
- Jev, a 'System One' model that only makes decisions, questions how many LLM calls agents really need — ThePromptIndex · 2026-09-20
- Bend 2 called a 'software factory language': formal verification ends code review — airesearch12 · 2026-09-20