How to Implement Visual AI in Complex Manufacturing

Brian Ton runs the quality labs at Florida Crystals Corporation as senior laboratory manager. He has implemented visual AI in a real manufacturing operation, with what he openly calls varied results, lots of mixed results.

That candor is what makes his recent appearance on the Emerj AI in Business podcast worth your time.

Florida Crystals farms more than 190,000 acres of sugarcane, rice and vegetables in South Florida, owns two sugar mills, a refinery, and a rice mill, and is the parent of ASR Group, the world’s largest refiner and marketer of cane sugar, whose brands include Domino, C&H, Tate & Lyle and Redpath.

Ton describes from inside a working plant, how to implement visual AI in complex manufacturing so it becomes something operators rely on rather than another pilot that gets shelved.

Here is what he told host Yolandi de Weerdt, and what you can take from it before your own first deployment.

The technology is not the hard part

De Weerdt opens on hidden quality gaps, the problems that became background noise, the things every shift has learned to work around. Ton reveals, “I think change management is one of the biggest universal challenges facing many businesses.”

His evidence comes from his own lab. Near infrared spectroscopy, a workhorse of his domain, was introduced somewhere between the 1960s and the 1980s. Days before the recording, he was at a conference where people were still presenting new ways to apply it. Technologies take decades to mature inside businesses, and AI has collapsed the window companies get to absorb each one. “Imagine looking through a telescope and then the lens of that telescope is constantly changing and the landscape that you’re looking at is constantly changing.” he says.

In short, an implementation plan is mostly a change management plan.

What he learned about replacing manual inspection

Asked where manual inspection breaks down, at volume, consistency, or attention span, Ton points at something more basic: “you’re dealing with the natural variability that gets introduced when you’re talking about humans.” Two inspectors will not judge borderline product the same way, and neither will one inspector at hour two and hour eight of a shift.

He also pushes back on the framing plant floors fear most, that the camera is there to replace people. “It’s taking manual labor and it’s turning it into a productivity multiplier. And at least that’s what I’ve seen in some of the limited use cases that we’ve successfully implemented.”

The three tests he applies before trusting a system

Ton also shares his checklist for whether a visual AI system will earn operational trust, meaning the people on the floor believe it will be there and be right when a shift changes or something goes wrong.

The first test is his number one: “Is there a robust system of validation and verification?” Whatever the tool, a vision model or an LLM, you have to be able to check its output against the pre-existing process it supplements. At Florida Crystals that means comparing the new system’s answers to the measurements the lab already trusts.

The second is accessibility across every level of the organization, operator to supervisor to middle management to directors, because those are different people with different relationships to technology. In Ton’s description, the same rollout has to work for someone who reads paper reports and for a hardcore vibe coder with agents running on Claude.

The third is proximity to subject matter experts. “The subject matter experts within the technical departments in your business are oftentimes not the programmer.” The people who know what good product looks like on your line usually did not build the model, and closing that distance, so the domain expert can shape and correct the system directly, is what he calls very advantageous. It is also the design principle behind quality control tooling that quality teams can own themselves instead of routing every change through a software team.

The thing he would tell his past self

De Weerdt asks whether a properly calibrated system can run as a static tool. “I very seriously doubt that anything can just exist as a static tool,” Ton says. There will always be oversights, something obvious to the person working the line that never occurred to the software architect.

The channel that catches those cases is the feedback loop: operators flag what the model got wrong, and the flagged cases go back into training. This is human-in-the-loop computer vision: “A feedback loop in terms of AI is almost like the quality control for the quality tool.” as Ton says.

When de Weerdt closes by asking what he wishes he had known before his first deployment, he does not hesitate. Design that feedback loop early, as part of the validation and verification cycle, not as a retrofit after the system is live.

Start with small victories

Why do organizations abandon visual AI before it matures? Ton’s answer is reach. “Small victories go a long way.” A home run project with a big technology gap produces a stream of small, discouraging surprises, and eventually someone hits the abort button.

His alternative: pick a problem conceptually close to home that everyone understands, solve that, prove value people can see, then scale. His bar for the first project is that the value be measurable, understandable, and achievable.

And the deciding factor, the thing that separates operations where visual AI becomes standard practice from the ones that stall? Not a technology choice. “It really comes down to persistence.” Promises do not move a plant. A system running reliably inside the operation does.

Hear the full conversation

The episode also covers how Ton thinks about static pressure in quality organizations and what shift-to-shift reliability demands of a system. Listen to it on Emerj. And as Brian signs off, “Good luck to everybody going down this path. It’ll be worth it in the end.”

Train a model on your own images and deploy it with Roboflow, or learn more about implementing Roboflow’s Vision AI Center of Excellence Blueprint.