Innovation Fall 2026

“

Sonibel Instruments, a Vancouver weld-defect detection startup launched by three University of BC graduates in 2025, is pursuing an alternative route to more adaptive welding. Its torch-mounted acoustic sensor listens to the arc and flags patterns associated with possible defects, giving a welder the opportunity to correct the process before a failed nondestructive test (NDT) leads to repair and reinspection. "What we're designing is to be able to give the welder the decision-making power, yet notify them of what we think is a defect or not based on data that we've gathered on high-mix processes," said Sonibel Instruments co-founder Hooman Pirouz. Vision and sound capture different information about the state of the weld. “It’s really a case of how you capture the data, how you train it, and how you get the sheer volume of data that’s required,” said Smith. Through a $1.8 million Mitacs grant, his research group has begun working with Seaspan and Simon Fraser University to pilot novel hardware and software approaches to building welding automation.

We'll see fewer humans physically holding the torch, which is the dangerous part because welding involves long-term hazards: exposure to arc light, carcinogenic gases, and repetitive strain injuries from doing the same motions for many years. Soroush Karimzadeh, P.Eng., Novarc CEO

”

to help Novarc move into general welding capacity, including structural and plate welding. For Novarc’s new NovAI product, the team seeks to add layers of software intelligence to existing robots rather than build new dedicated machines. Novarc has recently begun to enter discussions with clients and external manufacturers like Yaskawa, a global manufacturer of six-axis robots, looking to develop partnerships that package intelligence onto existing industrial machinery. "We place a machine-vision camera on the robot's end effector so it can see the weld pool," said Gumulia. "The camera sees the weld, the system makes inferences and recommendations, and those are fed back into the robot to control welding and motion parameters." The concept of using AI as a general intelligence layer for industrial machines remains early in its research and commercialization. Novarc's engineers are still refining how their models interpret the physical environments the robots operate in. "For example, shiny or polished parts can skew the vision model. The model must learn to handle those edge cases, which means being able to segment and interpret those images correctly," said Gumulia. "Identifying edge cases, seeing them in the field, and training our models on them is a big part of the day-to day work." Across BC, startups and institutions are testing different hypotheses as to what data inputs and processing methods will yield the most capable AI-enabled machines.

Innovation Fall 2026

29