Agents that write product copy and images for sellers, a shopping assistant that answers by voice, and the architecture that ties them into a store already running.
about 60%
less manual setup time
At LeanScale I designed how AI agents fit into an established e-commerce brand, and the system architecture for the whole integration. It has two audiences. Sellers get agents that take over the slow parts of listing products. Shoppers get an assistant that helps them find what they came for, by voice if they want.
A good listing needs a description that sells and pictures that show the product properly, for every product, and again whenever it changes. The seller agents rewrite descriptions automatically and generate new product images, so a seller reviews and approves instead of starting from a blank page.
On the other side, the discovery assistant helps shoppers find the product they actually want, not just the one that matches their keywords. It also listens: a voice pipeline with wake-word detection lets someone ask out loud, the way they would ask a person in a shop.
This is a store that is already running, so agents act on real products, real orders and other companies' systems. The setup agent shows the rule everything follows. It works in stages, and each stage hands the next something concrete that can be checked on its own.
Parse the intent
What is being asked for, in a structured form.
Select a tool
Which operation can do it.
Validate the action
Check the action before it runs.
Execute
Run it against the store or the third-party API.
A misunderstanding shows up as a wrong intent before any tool is chosen, and a right tool with wrong arguments is caught by validation. The model proposes and the code checks. With that in place, manual setup time for third-party integrations dropped by about 60%.
This was client work at LeanScale, so the code is not public.