I recently attended the Decorators Summit in Nashville by Printing United, and this was my presentation. °FAI: AI Pipeline for a Screen Print Factory:

Everybody wants an autonomous agent on day one. That’s exactly how you get bad output and torch your own confidence in the system before it’s earned any. So I don’t build generic tools. I build custom pipelines wired to one factory’s actual capability, prepress, generative art, proofing, QC gates, and every piece ships read-only. It doesn’t write anything without a human. People stay in the loop benchmarking, and only after it clears the accuracy gates you set, over time, do you decide whether to let it run looser. This is headed toward agentic. It is not agentic yet, and that’s the point. QC gate, before anything generates

A proof comes in looking fine. I upload it and the tool reads every field on the sheet against the artwork itself. Flag: the size field says the front is 13.75″ wide; the graphic placed on the sheet is 7″. Those two numbers contradict each other. Now I go back to the customer or fix it internally.

Generative art, trained on what already sells

I start from an existing artwork library, something established, something that already sells and build the generative tools around it, so what comes out matches the aesthetic you already own. I pull historical pieces and generate new work in that style. It can do a landmark swap (look up the location, pull something local, hold the same look), merge two images into one, or retarget the whole thing to a different city. One at a time or in batch, off a single image or many.

Proofing that matches what you already send

The proofing is custom too, it matches what you hand customers today, not some generic mockup. I pick from the shirt colors I actually carry, so the customer sees the art on the garment they’re ordering. Flat tee proof, illustration, photoreal, even a 3D turntable.

Vector sep

Then separation. The vector sep tool vectorizes the raster output, sets the color count, drops the background, runs the sep. It sets art to a preset size (8″ for a women’s cut here), one at a time or across a whole batch. Smart choking is built in, it handles gradients, distressed bitmaps, and knows PMS, your own custom palettes, branded apparel palettes, whatever you feed it. Every color lands on its own layer, selectable by color on the underbase. It’ll even re-center a graphic that’s technically centered in Illustrator but reads off to the eye, like a human would. It runs around 65 full seps an hour, generated AI graphic included or human made, and respects factory rules: one job gets an underbase, the other doesn’t, because the tool knows this factory doesn’t underbase that ink color on that shirt.

QC gate two — spelling, dates, name-and-number

The second gate checks the sep against the original if you want or it’s spellchecking a name-drop batch: it catches words that are misspelled and hands back a report naming exactly which graphics are wrong. It can verify online accuracy too, if a date’s tied to a graphic, it looks it up and confirms nothing’s changed or validates name-and-number sheets. These gates make API calls, so how much of your data ever touches a third party is the factory’s call: private model, frontier model, or hybrid. Your privacy, your cost, your choice.

Print order rotation: the advisor at the press

Finished sep, finished proof, now the press needs a rotation. This tool sits inside your ERP and pulls the job number, artwork name, location, garment color, brand, piece count, style, then looks up fabric content and shirt color, the size run image areas online and populates. The artist still enters ink system, line screen, and base; as it learns your methodology, it starts entering those too. Then it reads the separation and hands back a full rotation: where flashes go, what mesh, which flashes are recommended versus mandatory.

It can run a visual analysis to judge foreground, midground, background, and what’s actually dominant to the eye, then adjust color order, and you can toggle that off.

It’s an advisor, here to help a human make the call.

°FAI: AI Pipeline for a Screen Print Factory

What this really becomes – the sep tools feed the rotation tool, the rotation feeds the floor, and when something changes at the press, that change feeds back into the system — your inks, your templates, your underbase rules, your standards, getting more accurate on your own historical data. None of it replaces the printer’s judgment. It carries the judgment while you’re watching it. And eventually, when you’re ready, when you’re not watching it anymore.

Raster SEP is in Development now.

Tools used: custom QC gates, library-trained generative art tool, custom proofing, vector sep tool, print order rotation advisor. All built read-only, human-in-the-loop, benchmarked toward autonomy over time.