The framing on dresma.com is direct: "AI content creation and visual marketing" — competitor analysis in, AI mood boards out, conversion-focused visuals at the end. Every step is a model call: vision models reading competitor pages, language models scoring intent, image models generating, ranking models picking the winner.
That's a beautiful product. It's also a stack of vendor invoices that grows every quarter as new model providers ship better, cheaper, or faster options. The hardest engineering problem isn't getting any one model to work — it's the runtime between Dresma's code and twelve different model APIs. That's where Cloudflare's most timely thing slots in.
What that means specifically for Dresma:
Cost predictability — every customer's monthly AI budget becomes a hard ceiling, not a hope.
Provider portability — when image-gen-V4 ships next month at a third of the price, you swap the route, not the integration.
One observability surface — token spend, latency, failure rate, and prompt-level cost attribution across every model, in one log.
Resilience by default — provider outage cascades to fallback automatically; the four-step workflow doesn't stall when one vendor has a bad day.
As the model roster grows from 4 to 14, is the harder problem the provider sprawl (integrations, retries, fallbacks) — or what the CFO sees when the AWS plus model-provider bills arrive together? 20 minutes to compare notes on what we're seeing across AI-native customers.
AI Gateway is one piece. The full version maps Dresma's Track → Curate → Create → Human QC workflow to Cloudflare primitives — R2 for the asset library (zero egress vs. S3 plus CloudFront), Vectorize for mood-board similarity search, Images and Stream for output optimization, and a worked AWS-vs-Cloudflare cost model for an asset library at scale.
Read the architecture walkthrough →