A buyer ordering 300 branded fans for a summer event rarely knows the product is filed as a "mini handheld fan with phone stand." They have a photo from a competitor’s flyer, or a picture in their head, and a search box that only matches the exact words in a product title. On a catalogue of more than 100,000 items, a wrong word returns nothing, and a buyer who finds nothing leaves.
The network made the problem bigger. More than 2,000 franchise storefronts sell from the same catalogue, each with its own address, its own products switched off and its own prices. Any new search had to work on all of them at once, without each franchise owner touching code, and without the head office running two thousand copies of anything.
The search is a separate service that sits beside the existing site rather than inside it, so nothing in the old shop had to be rebuilt.
Search by photo, by words, or both. Every product’s picture and description are turned into vectors by an embedding model and stored in one search index. A buyer can upload or paste a photo, type a plain description, or combine the two, and the closest products come back ranked by similarity.
Three answers, not a wall of results. Each search returns the three closest products, a longer list of related ones, and the category the buyer is probably looking for, so a near miss still leads somewhere.
One index, every storefront. All franchise sites share the same index. A small widget adds a camera button beside each site’s existing search box, and each storefront’s own rules are applied at query time: products it has switched off are hidden, its own prices are shown, and links go to its own pages.
In step with the catalogue. A sync worker reads the live product database read-only and re-indexes only what changed, so a product taken off sale disappears from search without a full rebuild. Only items on sale and on display are ever returned.
Built to stay up. Results are cached, uploads are checked for format and size before they are processed, and every storefront and visitor is rate-limited, so one heavy user cannot slow the search for everyone else. Searches and clicks are logged, and a quality report shows which queries find nothing.
A buyer can now search with what they have. A photo of a product, a description in their own words, or both, returns the closest items in the catalogue, with how close each one is.
The whole network got it at once. The head-office site and more than 2,000 franchise storefronts run the same search from one index of 103,632 products, each with its own exclusions and prices, and none of them had to change their shop to get it.
The catalogue and the search no longer drift apart. Changes in the product database reach search through incremental sync, so a discontinued item does not keep turning up in results.
The search is in daily use across the network. The next step is using its search and click logs to tune ranking for each kind of storefront.
Rian Blue reads a mechanical bid set the way an estimator would, finding every duct run, size, fitting and air device, and turns it into the estimator’s own workbook. The estimator answers only the questions the tool could not settle on its own.
A marketing team running accounts on six platforms was writing every post six times. We built one workspace where a post is written once, rewritten for each platform from the team’s own past posts, scheduled, and measured in the same place.
Patients with depression see their psychiatrist for minutes at a time, and the weeks in between go unrecorded. Sogon-Sogon is a voice-first AI journal that turns a patient’s day into a diary entry for them and a report their clinician can read.
