Fariya Raza
System 02

Creative Generation Pipeline

On-brand ad creatives at volume, QA-checked by a vision model. 100+ creatives a day for an EU education brand, each checked against brand rules before a human sees it, at 78% lower inference cost.

100+creatives a day, QA-checked
Hover a step for what it does

Text version: Brief → Prompt → Generate → Vision QA → Regenerate → Approve

Problem

The brand needed a steady stream of Meta ad creatives without a designer reviewing every image.

What I built

A scheduled n8n pipeline: brief → cached prompt → image generation → vision-model QA against written brand rules → automatic regeneration on failure, capped.

Key engineering decisions

  • Prompt caching on the stable brand context cut per-image inference cost by 78%.
  • A vision model reviews every output against explicit rules (logo placement, palette, text density) rather than a human eyeballing batches.
  • Capped retries: a creative that fails twice is parked for review instead of looping.

Validation and QA

Every creative carries its QA verdict and the rule it passed or failed; the pipeline logs all runs.

Result

100+ QA-checked creatives a day, running on schedule whether anyone is watching or not.