Coco_vSp
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Website-care teams handle a constant stream of small change requests from customers, such as “our phone number changed, please update the header and footer.” A builder makes the edit, says it’s done, and the request is closed. Often it isn’t really done. The footer still shows the old number, the mobile menu wasn’t updated, or a link points to the wrong page. The customer finds the mistake, the work gets redone, and trust erodes. No one checks every affected page at every screen size by hand, because it doesn’t scale. Feluda closes that gap. It reads the customer’s request, crawls the site, and uses Muse Spark 1.3 to split the request into precise jobs, one for every place the change has to appear. A Care lead approves the plan, Feluda captures the site’s “before” state, and each builder gets validated, step-by-step editing instructions. When the builder clicks Mark done, Feluda checks the live site itself. Playwright captures every affected page at desktop and mobile widths. DOM checks confirm the exact text and link, and that the old value is gone. SAM 3.1 locates the changed element in the screenshot, and a Muse Spark visual judge compares before and after views against the request. Each job is passed, or failed with a specific reason the builder can act on. A request can only close when every job passes and a final check of the whole site confirms nothing else changed. Muse Spark then drafts the customer’s “done” note for a person to review and send. In our demo, a builder fixes the header but forgets the footer. Feluda passes the header, fails the footer with the exact reason, and blocks closure until it’s fixed. A full request cost about $0.05 in model spend, and each verification took about 15 seconds. Built with Meta Muse Spark, Muse Wonder, SAM 3.1 and Muse Image, plus FastAPI and Playwright.