AI startup Groundlight has launched a visual understanding service based on natural language instructions. The aim is to accelerate development of computer vision applications for manufacturing, industrial automation, process monitoring and robotics. The service enables a developer to understand images programmatically using simple English language instructions and just a few lines of code.

Groundlight’s platform is designed to allow developers to quickly build robust vision solutions for industrial purposes. It is intended to replace the need to gather and label a dataset, train a model, and then develop an MLOps solution to maintain it. Visual tasks can be described in minutes in natural language, which is then turned immediately into an application-specific model. Groundlight tracks the confidence of this model and continuously optimises it using feedback from expert human monitors.

“As the most powerful AI models get better, there is the tendency to think that all you need is a Large Language Model. While they make great demos, robust commercial systems need real human judgement to handle edge cases reliably. And massive LLMs are too slow and expensive to be relied on in isolation for many applications,” said founder Leo Dirac. “Groundlight’s approach seamlessly ties together traditional deep learning with massive foundation models, edge computing, and real-time expert human supervision.”

Groundlight says it can enable small and mid-sized manufacturers to rapidly develop ML solutions that increase productivity. Austere Manufacturing in Washington State produces cam buckles. “Quality control, process efficiency, and continuous improvement are crucial to our success. We’re excited to use Groundlight to inspect our products and monitor our processes without the development overhead of a typical industrial solution. Their API enables a $10 camera and a few lines of code to implement a working solution in minutes. Far less time than we’d spend to even evaluate an expensive industrial computer vision product,” said Uriel Eisen, Founder of Austere Manufacturing.

Groundlight is also applicable to companies with large retail and warehouse footprints. Even if these companies have existing solutions, implementing the same ML solution in a new environment often means starting from scratch on data collection and training. Groundlight can shorten the process of re-implementing existing AI models to new environments.

“So many ML projects don’t even get off the ground because of the big effort and scarce expertise required to credibly investigate the feasibility of a solution, much less implement it. Then they require operational expertise to keep them running. Groundlight has the potential to change all of that,” said Thomas Stubbs, ARC Labs, Slalom Consulting.

“The newest generation of foundation models are incredibly powerful, but they are both expensive to operate and unreliable for many use cases. Every company has unique data, especially in industrial applications. Groundlight’s approach enables companies to utilise off-the-shelf cameras and inexpensive equipment to quickly build and reliably operate customised models. We are excited to work with this brilliant team as they build the company,” said Tim Porter, managing director, Madrona.

Co-founded in 2019 by Amazon AI veteran Leo Dirac and Microsoft Hardware alum Avi Geiger, Groundlight is emerging from stealth, backed by $10 million in new funding led by Madrona, with participation from Greycroft Partners, Founders Co-op, Flying Fish, AscendVC, and EssenceVC.