E-commerce
New product launches built in under an hour, with a person approving before anything goes live
Raghav’s work at a Canadian sports memorabilia retailer, where he works now.
Three timed launches went from request to built in 13, 24 and 34 minutes, with inventory problems caught before customers saw them.
- From request to built, across three timed launches
- 13–34 min
- Shopify tools the agent works with, each checked or read back
- 44
- Missing products traced to one NetSuite setting before launch
- 17
The situation
The retailer launches limited-edition products around athlete signings and events, often on short notice and timed to an email campaign. Each launch meant hand-building products, variants and theme pages in Shopify, while inventory lived in NetSuite.
What was hard
Speed creates mistakes, and mistakes on a live store cost sales and trust. The two systems also fight each other: NetSuite overwrites Shopify inventory on every sync, so a stock fix made in Shopify silently disappears.
What Raghav did
- Built an "AI operator": a Python server with 44 tools that lets AI agents (Claude and Codex) work directly in the Shopify admin through its GraphQL API.
- Designed the safeguards in. Theme changes are dry runs until explicitly applied and are backed up first, new products start as drafts, invalid theme files are refused, and every change is read back to confirm it landed.
- Had the agent cross-check NetSuite, read-only, before each launch.
- Set the rule that launches are built on a test theme and go live only on an explicit go-ahead.
What changed
Three timed launches went from request to built in 13, 24 and 34 minutes; one went live about an hour before its email blast.
Before one launch, the agent found that 11 of 12 products existed in NetSuite but none were in Shopify, and that 6 had no stock. It traced 17 more missing products to a single NetSuite setting. Because it understood the sync, it flagged the stock problems instead of "fixing" them where the fix would be overwritten.
How this was measured. Times run from the request to the built product, taken from the timestamps of the working sessions.
What went wrong
Once, the agent published 8 draft products without review. They were live for about 10 minutes, received no orders, and were rolled back. The rule now: the agent builds on a test theme, a person says go, and every write is read back.
Bring us in when your team launches products on a deadline and you want speed without gambling on your live store.