Shelfmark (Pittsburgh, PA, USA) recently completed a funding effort that has raised $3.5 million in seed money to further expand deployments of its AI-based manufacturing intelligence platform. The company, a start-up that specializes in physical AI solutions for manufacturing, developed a holistic manufacturing intelligence platform that uses in-line industrial cameras, spatial sensing, and proprietary deep learning models designed to perform automated inspection and data interpretation, in real time, in fast moving manufacturing settings such as reels, rolls, and continuous web manufacturing facilities.
Products such as industrial films, graphics, webbing, paper, flooring, and coiled metals are made on fast moving lines, and often the process involves frequent specification changes, which make visual inspection and defect detection difficult, especially in real time, notes Shelfmark CEO Pat O’Donnell.
“Too often, issues are discovered after large volumes of affected material have already been produced,” he says.
Shelfmark launched in 2022 with the initial goal of bringing AI-based automation to companies in western Pennsylvania that are underserved by automation processes. The company has worked with more than 40 area companies and has helped them achieve 99.5% defection detection accuracy.
“We built Shelfmark in Pittsburgh, alongside operators and engineers on real factory floors, to give those lines the intelligence they need to become more autonomous,” O’Donnell said. “This isn't about replacing workers; it's about doing work people can't perform consistently at line speed, catching problems as they happen and giving teams what they need to prevent problems in the future.”
With the latest round of fundraising, the company plans to expand its footprint, both geographically and in industrial sectors.
Vision Systems Design was interested in finding out more about the platform, so we reached out to O’Donnell for more information.
Editor's note: the following Q&A may have been edited for style and/or clarity.
Vision Systems Design (VSD): Continuous web inspection has been around for a while. With that in mind, what technical advances have enabled Shelfmark to achieve 99.5% defect-detection accuracy in environments where traditional systems often struggle?
Pat O’Donnell (POD): The key is the core deep learning (AI) technology that allows for defect detection in highly variable or complex products, which is just about anything made on a web. Shelfmark’s technology allows customers to detect defects without a recipe and dramatically improves upon rules-based approaches that have historically prevailed.
VSD: How does Shelfmark’s deep-learning approach handle new defect types, material variations, and production changes without requiring extensive retraining or manual tuning?
POD: Shelfmark’s combination of proprietary hardware bundles and deep-learning approach allows for novel flexibility handling material and product variation. While we can’t go too deep into the tech, the deep-learning algorithms are key to handling the variety.
VSD: Please explain the role of spatial sensing in your platform. How does it complement traditional line-scan or area-scan camera systems?
POD: Spatial sensing is a broad term for the environmental monitoring that allows the AI to correlate defects with the conditions that created them. This may be temperature, pressure, humidity, or other factors unique to the particular production method that we are monitoring.
VSD: Many manufacturers already have inspection cameras installed. How does Shelfmark integrate with existing vision infrastructure, PLCs, and MES systems?
POD: Shelfmark has a PLC integration layer that natively integrates with existing hardware. I would, however, challenge the notion that most manufacturers successfully use vision today. The customers we work with struggled with legacy vision systems and are eager to explore new technology that is more resilient and requires less expertise and management.
VSD: You emphasize causal AI rather than simple correlation. How does the platform distinguish between factors that coincide with defects vs. those that are actually causing defects?
POD: Again, without getting too deep into the intelligence layer itself, the key is building datasets that have enough context beyond simple binary good/not good. Systems today are not equipped to understand the entirety of the production environment and then fall short on interpreting the data that is collected. This is exactly our focus.
VSD: What are the biggest technical challenges in moving from defect detection to closed-loop process optimization?
POD: Closed-loop optimization is the billion-dollar solution that will require significant investment—in integration, root cause analysis, materials handling, and beyond. We don’t claim to be close to closed-loop optimization but are excited to play a small part to work towards this larger automation.
Related: From Camera to Insight: Platform Simplifies Industrial AI Inspection
