Half a million assets, and nobody could find anything
Eleven teams shared one media library, and searching it had become slower than shooting something new.
I built the AI layer over it: a model proposes metadata across 34 fields, a reviewer approves or corrects it in a console of my own, and every correction feeds the next run. A field writes back unattended only once its confidence clears 90 to 95%, which is what let it run unsupervised.
- Time per asset
- 3 to 7 secondsfrom 5 to 30 seconds by hand
- The library
- 500,000 assetsyears of shoots in one place
- Groundwork
- 80,000 imagestagged by hand first, which set the rules
iconik DAM · vision and language models · human-in-the-loop console · confidence thresholds

