AI agents & scheduling

An AI agent ran a carton printing press for a small SoCal manufacturer

Brian Roemmele reports wiring an AI agent into an existing flexographic carton-box printing line, with small local models watching the machine and cameras feeding it visual data.

Evidence: Self-reported by the builder on X. Early, no independent verification yet.
Source date: 2026-10-10 · Checked 2026-10-23 · Tech: AI agent, local AI models, video cameras

The problem

Short, custom print runs are expensive on a carton press: every changeover wastes material and time, and combining several customers' jobs efficiently is a puzzle people solve by hand.

What they did

According to his posts, the agent was given access to the same screens an operator uses (machine HMI, job management, MES/ERP and camera software). Six local AI models each watched part of the machine and reported to the agent, which planned and sequenced the print runs. He says it was set up for a long-term client in Southern California after several test runs.

Reported results

He reports a 3-hour supervised run with no problems, and says re-arranging the print runs cut paper and ink use by 58% and compressed about four days of scheduled work into three hours. Before the run he projected up to 30% higher machine utilization.

Our take

This is the most ambitious claim in our collection and the least verified: it comes from the builder himself, a few days after install, with no published video or customer statement yet. The interesting, transferable idea is smaller than "AI runs the factory": job combining and run sequencing is a classic scheduling problem, and an assistant that can read your existing software screens may help without replacing your control system. We'll update this page if he publishes follow-ups.

Try this on Monday: Before handing anything to an agent, export a week of jobs and ask whether better batching alone (same material, same setup) would save changeovers. That part needs no AI at all.

Sources

Found via @BrianRoemmele on X. Summarized in our own words; see the original posts for full context.

More case studies

Vision on a budget

A Raspberry Pi and a camera running a vision model with no cloud

Brian Roemmele showed a Raspberry Pi with an attached camera running the open LLaVA 1.5 model locally to describe and recognize what passed in front of it.

Evidence: Hobbyist demo shared on X (2023). Not a factory deployment. · via @BrianRoemmele