I Stopped Asking Gemini 'Is This Right?' and Started Hiding My Answer Instead
Ask Gemini to confirm a label you already chose and it tends to agree. Here is why I compared three ways of asking, chose blind re-classification, how my answer still leaked in, and the routing code that sends only disagreements to a human.
An agreement-rate gate approves the model swap, then your category shares move
Swapping the model behind an image classification batch can pass a golden-set agreement gate while category shares quietly shift. I measured the sample sizes each check really needs and rebuilt the gate as a paired comparison.
A near-miss label won't fix itself on retry — a normalization layer for closed-vocabulary classification
When responseSchema enum returns an out-of-set label, retrying tends to return the same near-miss. From a wallpaper app's 30-category batch, here is the distribution of how labels miss, plus a normalization layer built on an alias table and gemini-embedding-2 nearest-neighbor, with measured results.
The Day We Went From 30 Categories to 34 — Reclassifying 1,180 Assets Instead of 8,142
Adding categories to a taxonomy does not require reclassifying everything. Here is how embeddings and confidence margins narrowed a backfill from 8,142 assets to 1,180, with the numbers.
Auto-Categorizing 3,000 Wallpaper Images With Gemini Vision API — A Real Production Account
Manually categorizing thousands of wallpaper images doesn't scale. This is a hands-on account of building an auto-classification pipeline with Gemini Vision API — covering design, implementation, actual cost, and the failure patterns I hit running 3,000 images through it.