AI Creative Pipeline for Growth

At Bestie, growth ads are generated by a multi-agent AI pipeline, not drawn by designers. My deliverable was no longer the interface. It was the rules the AI must follow.

A product designer by title, I stepped in when the growth team's pipeline started shipping off-brand at scale: I owned the brand-constraint layer, and contributed to the asset taxonomy and the human-in-the-loop review tools that keep AI-generated ads on-brand.

Nine specialised agents cover the creative lifecycle: one diagnoses published ads, three generate static images (iterating winners, proposing new concepts, remaking underperformers), and five more run the video studio chain. I own the visual rule layer that governs all of them, and designed parts of the operator tools around the loop. The agent architecture itself is engineering-led.

The static-image path, where the rule layer does most of its work. The five video agents run a parallel chain under the same constraints.

The loop the diagram abstracts: each generation of variants is scored, and the system keeps the winners and kills the rest. Figures redacted.

A pipeline can only enforce rules; a human still has to decide what good creative is. Grounded in the team's user research and competitive analysis, I translated who the user actually is into the ad concepts the pipeline generates, and into what “on-brand” means for it to protect.

Every generated ad has to survive this translation: real emotional need in, on-brand creative out. That standard is exactly what the guardrails below defend.

Left unconstrained, the pipeline drifted: character avatars were misused as logos, and product UI was invented by the model instead of being sourced from the real asset library. Each drifted ad quietly eroded brand trust, at generation speed.

The scale problem wasn't one bad ad, it was compounding: the pipeline generates on a daily automated schedule, so a bad rule reproduces its mistake every single morning until someone codifies the fix. I flagged the two failure classes and defined what “correct” looks like; the fix below shipped as a pair with engineering.

The audience of a design spec changed: it is now read by a model, not a person. I led the fix, translating visual rules into prompt-level hard constraints, turning “the avatar is never a logo” from a guideline humans agree with into an instruction the model cannot ignore (shipped with engineering).

I contributed to a four-category use_case taxonomy for the asset library, so the pipeline could caption and compose assets according to ad context: information architecture in service of model behaviour.

A pipeline is only as reliable as the tools operators use to steer it. I contributed to the design of the operator toolchain: an asset manager with auto-captioning, a prompt editor with dry-run validation, an ad review queue feeding Meta publishing, and a competitor ad browser.

The context these rules protected. Scaling paid growth is a whole team's work, and the numbers below are theirs, not mine — my job was to keep that volume on-brand as it grew. Over seven months monthly ad spend scaled roughly 8x while Meta CPI fell 46% and blended CPA fell 32%.

The rules were accountable to real numbers: this is the analytics surface the growth team steered by. Values redacted.