AI Image Generation Make Workflow

Redesigned and rebuilt the image generation module for an AI-powered course automation system serving Naturalistico, one of the largest holistic certification providers globally (300K+ students, 11 brands, ~5 courses/month).

The existing Make.com pipeline blindly inserted 2 AI-generated images per subsection (~200 images/course), resulting in $200-300/month in OpenAI image costs, high editor rejection rates, and zero diagram support despite course content heavily benefiting from it.

What I Built

Image Placement Analyst — an LLM module that reads the full edited HTML of each course module and decides where images should go, what type (stock photo / AI illustration / diagram / skip), capping at 1 per subsection and 4 per module

Model migration from GPT Image to Nano Banana 2 (gemini-3.1-flash-image-preview) with Batch API for 50% cost reduction

Three-track parallel router splitting jobs to Pexels (stock), Nano Banana 2 (illustrations), and Napkin AI (diagrams)

Vision-based quality validator using NB2 to reject off-topic images before AWS upload, with 1-retry loop

Diagram generation via structured NB2 prompts for concept maps, cycle diagrams, and infographic-style clinical frameworks