Case Study - An app that scores a public website’s design, then turns the low scores into a call list.
A marketing founder wanted a list instead of an afternoon of clicking through directories. We built an application that finds businesses, captures the public page, and asks several vision models to score the design. Low scores become leads.
- Client
- Marketing studio
- Year
- Service
- Full-stack, vision models
Overview
Agencies still prospect by opening a directory and deciding, by eye, which public site looks tired enough to call. A marketing founder wanted that loop to be a list, not a hunch.
We built the application from an empty repo. A user names a place and a trade. The system finds businesses, captures the public page, and asks vision models to score the design from 1 to 100. Low scores can be filtered, enriched, and exported into the outreach tool the user already has.
Design quality is a judgment, and no single model was steady on every page. We routed the same captures through several model families, compared them on consistency and cost, and kept tuning the prompt until the score stopped jumping between runs. Layout, type, color, and whether the page looked finished were the axes. The shipped scorer blends the families that held up.
The app is a modern web stack: a hosted database, sign-in, a cache in front of search, object storage for the captures, and a background job for the crawl-and-score loop. Payments and the export path landed in the same four months as the model work.
What we did
- Full-stack app
- Vision scoring
- Several model families
- Place search
- Lead export
- Model families compared
- 4
- Empty repo to production
- 4 months
- Design score
- 1–100