Captured from the live deployment, not a mockup
01 · The problem
A blurry photo of the right bumper, a sharp photo of the wrong bumper and a sharp photo of real damage look identical to a naive classifier. Built a four-stage vision-language pipeline that separates them.
02 · How it works
- 01
Analysed each image in its own call rather than batching, keeping every visual judgement independent, and defended against prompt injection arriving as rendered text inside the submitted photos.
03 · What it cost, and what it returned
Kept the risk engine pure Python with zero model calls so aggregation cannot be talked out of a verdict, reached 75.0% on the headline decision, and published that n=20 makes every figure low-resolution.