Large features built on a live platform, at a fraction of a rebuild.
This is a large WordPress fine art marketplace, more than 120 partner galleries and around 26,000 works, that had passed through a succession of external developers over the years. Several core parts had started working against each other: the gallery directory was too limited, artwork filtering was inconsistent across templates, and legacy image handling was creating avoidable drag.
The symptoms were visible to anyone using the site. Gallery preview pages took around seven seconds to generate, firing more than 2,100 database queries each. The site felt slow, and the underlying causes were opaque.
The project ran on three fronts. On discovery, I turned a simple gallery listing into a proper directory product: filtering by artist, country, art movement and gallery, clean shareable filter URLs, featured artists, dynamic headings, and routing straight into pre-filtered pages. I rationalised filtering across every template, fixed a bug that was wrongly including sold works in "Buy it Now", and added two new currencies across the whole artwork stack.
On stability, I found the cron was not rescanning artworks at all. It was endlessly retrying a stuck queue of 27,780 problematic attachments, hour after hour, achieving nothing. And on diagnosis, I measured properly: capturing page-generation time, query counts and memory alongside server load to prove the bottleneck was the application, not the server, so the money went into fixing code rather than buying hardware.
The core of this was delivered in around two months of actual development time, using AI-assisted development, taking over from years of previous developers.
The marketplace now has a materially stronger discovery layer for galleries and artworks, more accurate commerce filtering, two new currencies, and a cleaner template setup that is manageable through WordPress as intended. The large features were delivered on the existing platform rather than through an expensive full rebuild.
Just as valuable is the clarity. A 27,780-item image backlog identified and characterised, the seven-second requests pinned to application behaviour rather than infrastructure, and several previously opaque problems turned into measurable, fixable systems.
Refusing to throw hardware at a software problem. The rigour to measure before acting, proving the server was fine and the code was not, meant the client spent on the actual cause.
And using AI-assisted development to build substantial features quickly, on foundations that were fixed rather than replaced. That is how two months of work replaced what a rebuild would have cost.