We Raised $3.5 Million. Here’s Why.

Quality is in the air

Early on, we worked with a plant where the defect rate seemed to change with the weather. Nobody could explain why. It was the same line, material and crew, but some weeks everything ran clean and other weeks it didn’t.

Our AI visual inspection system was doing exactly what we’d designed it to do: watching every inch of material at full speed and flagging defects in real time. What we hadn’t realized was that it was also building a record of the line, capturing both the defects and the conditions surrounding them.

That record held the answer. When temperature and humidity fluctuated, the defect rate moved with them. The plant installed humidity controls, and the defect rate fell by half.

The fix was a fact about the air, sitting in the data. It also changed how we thought about the problem. We had started by helping the plant see defects it couldn’t see before. Now, the system was helping explain why those defects happened, and how to stop them from happening again.

What we set out to fix

Shelfmark started with a $5,000 check on a factory floor an hour outside Pittsburgh. We were there trying to sell AI-powered inventory cameras. The key word is trying.

As we explained the technology, the plant manager said, “That’s great and all, but can you point those cameras at my printers and tell me when something goes wrong?” Then, almost as a joke, he added, “I pay people to stare at them all day when they could be doing better things. If you can fix that, I’d hand you a check.”

His problem wasn’t his alone. It’s one of the defining quality challenges in high-mix, continuous-flow manufacturing across decorated apparel, industrial films, labels, textiles, coatings and more. Material moves too quickly for people to inspect consistently, while defects can look completely different from one product to another. Traditional rules-based machine vision systems often struggle with that much variation.

So, inspection stays manual. In some plants, people spend entire shifts watching material move faster than the human eye can properly inspect. For one customer, inspection meant unrolling material as long as a football field and checking it foot by foot. That was how the job was done 40 (or 400) years ago, and on many lines it’s still done that way today.

These aren’t manufacturers that have failed to modernize. They are sophisticated businesses making technically demanding products for customers with extremely high standards. The problem is that technology built for highly standardized production environments hasn’t worked well on their lines.

We built a better answer: in-line cameras and deep-learning AI models that detect defects without requiring a separate reference image for every SKU. Once the system is mature, it inspects every inch of every run at production speed. We manage the hardware, models, tuning and upkeep, so customers never have to build or maintain a rule.

It took building alongside 40 manufacturing facilities to get here. We couldn’t develop the system in a lab and drop it onto the floor. We had to learn how light behaves on reflective films, how defects vary across materials, what matters to operators and how a system survives in a real production environment.

Today, Shelfmark can detect defects with up to 99.5% accuracy, and some customers have cut waste by as much as 90%. We were right about the first problem: nobody can reliably watch every inch of a fast-moving line, and nobody should have to.

What we found underneath

Years on production floors taught us that catching a defect is only the first step. A problem appears and someone catches it, whether that’s our camera or a sharp-eyed operator. The crew adjusts the process, the run recovers and everyone moves on. Then, a shift or material change later, the defect returns and the hunt begins again, as if the line has never seen it before.

An experienced operator learns. After 20 years on one line, they can sometimes hear a problem coming. They know which supplier’s rolls tend to run dirty, which settings need watching and what a humid week does to the scrap rate. They carry years of connections between what happened, what changed and what finally fixed it.

The industry often looks at that person and worries about what will happen when they retire. We took a slightly different lesson: the line is still running, and it holds much of the same information. You just have to listen closely enough.

Every run leaves a trail: camera frames, sample pulls, quality logs, material batches, machine settings, operator adjustments, temperature and humidity. But that production data is scattered across systems, spreadsheets and people. By the next shift, the connections are already disappearing.

Most vision systems see the problem and then forget the lesson. We want to give the line a memory: one record connecting every defect with the conditions around it, run after run.

That’s what found the answer at the plant that was impacted by humidity. It’s also what we mean when we say the line should learn. Each run should leave the next one better informed, so manufacturers don’t just detect defects, they understand their root causes, prevent recurrence and steadily improve how the line performs.

What the $3.5 million is for

The round was led by Armory Square Ventures, with participation from Grand Ventures, Hyde Park Angels, Argon Ventures and Cultivation Capital. It brings Shelfmark’s total funding to about $5 million.

The money is for the second half of the job. Today, Shelfmark can see every inch, catch the defect and connect it to what was happening on the floor. This round helps us move from explaining the last defect to predicting the next one and, ultimately, preventing it.

The press release calls this Physical AI for continuous-flow manufacturing. On the floor, it means something simpler: the line sees the problem, understands what caused it and stops making the same mistake. Over time, every run makes the line smarter.

This isn’t about replacing workers. Spending an entire shift staring at material moving faster than the human eye can inspect isn’t a good use of anyone’s time or talent. Our job is to take on the task no person can do consistently at line speed and give the crew information they can use to make better decisions.

The experienced operator will always understand the line in ways no camera can. We aren’t trying to replace that judgment; we’re trying to strengthen it, with a system that sees everything, understands and remembers what happened and helps the line keep up with the people who know it best.

The demand is already here. Across the four industries where we started, nine out of ten pilots have converted into customers. For them, this round means fewer bad rolls leaving the plant, less material in the scrap pile and a quality story they can prove to buyers. It also means their people can spend less time staring at material and more time doing work that calls on their judgment and experience.

Our vision hasn’t changed since that first $5,000 check: managed frontier AI on every continuous production line in America. It started with a plant manager joking about paying people to stare at printers when they could be working on better things. This round is the same bet at a different scale: that somebody should own this problem completely, solve it properly and stand behind the result.

This is no longer chapter one for Shelfmark. We know exactly what we’re building: frontier AI for the lines that build for the world. We’ve helped those lines see. Now we’re helping them learn.