AI Engineer — Automated Product Image Pipeline with Claude Code (Amazon, Nano Banana / Gemini)

We're building an automated pipeline that generates Amazon-ready listing images at scale — main images, infographic A+ content panels, and lifestyle shots — from a product photo plus its listing data. The system uses Google's Gemini image models (Nano Banana 2) with an AI design layer that writes the prompts and renders a full infographic grid in one pass. It works well on many products. The problem: output variance is too high. Some runs come out great, others come out broken (wrong product shape, melted details, garbled text) — there's nothing in between. Our goal is a hands-off system reliable enough to run on 10,000+ SKUs with only a small manual review queue. We need help with: Reducing run-to-run variance in image-model output (best-of-N selection, temperature/sampling control, fidelity gating) Building a reliable automated "is this the correct product?" check (vision-based fact gate, not taste) Hardening a Python + prompt pipeline for batch reliability at scale You should have: Strong Python + hands-on experience with generative image models (Gemini / nano-banana, or comparable diffusion/multimodal models) Experience taming non-determinism in LLM/image pipelines Bonus: e-commerce / Amazon listing image experience Send a short note on how you'd approach cutting variance in a stochastic image pipeline, plus one relevant project.