Agentic automations for Meta Muse Spark, and a generative video ad studio

Meta's AI Pathfinding team in Meta Superintelligence Labs wanted Muse Spark to run automations inside Meta's own products. Every automation had to fit how the owning product team already worked. And the output had to be good enough to ship to real Instagram and Facebook users, which is a much higher bar than a demo.

We prototype and build automations that Muse Spark can run inside Meta products. One of them generates AI ad concepts. We prototyped it and shipped it on Instagram and Facebook.

We also led a 0 to 1 build of an AI-native ad platform. It makes short-form video ads personalised to each viewer, and every impression renders its own video through Veo / Sora-class models, so there is no pre-produced creative at all. On top sits a conversational editor for marketers. They ask for edits in plain language, like "make the hat red" or "tighten the cut at 0:03". Edits stack across frames and takes without overwriting earlier changes, so nobody scrubs a timeline or starts a prompt over.

The core is agentic workflows built on tool calling, written in Python and TypeScript, with Node.js services and API design for the ad studio.

The rest is iteration. We tune prompts and model behaviour until the output is ready to ship, and we work directly with the product team that owns each surface so the automation fits how their product already works.

We shipped AI-generated ad concepts on Instagram and Facebook. We took the generative video ad studio from inception to launch, including a conversational editor that lets marketers change an ad by describing the edit.