The investment enables LatticeFlow to expand the capabilities of its platform and respond to growing customer demand as more companies and AI scaleups look to deploy computer vision models at scale.
The investment in the ETH Zurich spin-off, which brings total funding to date to $14.8m, was led by Atlantic Bridge and OpenOcean, with participation from FPV Ventures and existing investors btov Partners and Global Founders Capital.
The potential addressable market for computer vision is sizable due to its rapid adoption rate from manufacturing, healthcare, retail, security, and safety industries that are digitizing processes to become more data-driven.
In the past few years, computer vision AI models have surpassed human-level performance across image classification, detection, and other tasks in the lab. However, models often fail to work as expected when deployed in production because real-world scenarios are far more complex and varied than lab training datasets. Because of this 90% of all models don’t reach production, resulting in billions of losses.
“LatticeFlow is an enabling technology that empowers engineers and companies to deliver quality data and performant computer vision models that work in the real world,” said Petar Tsankov, Co-founder and CEO, LatticeFlow. “As data and models grow, delivering AI models that work in the wild becomes an unwinnable battle, so we built the first smart platform that empowers engineers to accomplish this task, addressing a major pain point.”
The LatticeFlow platform was built to automate the process of solving data quality and blind spot issues in computer vision AI models, critical to enabling model performance in the wild.
Data issues: LatticeFlow is unique in its ability to automatically discover and fix data quality issues at scale across datasets of millions of images, including labelling errors, poor-quality samples, data biases, and others.
Model blind spots: The platform also automates the discovery of blind spot scenarios, often impossible to spot manually, and fixes them before real-world performance is impacted. To patch the model, LatticeFlow has developed a new, scalable method for targeted data augmentation.
