According to US-based Grainge AI, food manufacturers are making critical production decisions based on outdated and incorrect metrics. Grainge AI is a startup developing machine learning tools that help food companies identify and predict the behavior of ingredients in their formulations.
Grainge AI was founded 18 months ago by food chemistry and AI researchers at the University of California, Davis, and builds software that determines the best testing protocols for ingredient applications. The company is targeting what co-founder and CEO Tarini Naravane calls “measurement blind spots” across the food industry.
“You can measure 10,000 things, but which ones are important to you?” Naravane says Ingredients first. “That's the problem we solve. The big question everyone is asking is, 'What data should we measure?'”
Co-founder and CTO Gabriel Simmons says this problem is pervasive in the food manufacturing industry. “People are making important decisions, for example, setting process parameters based on metrics that are 50 to 100 years old, such as how much protein is in this grain,” he explains.
“More detailed measurement techniques are available, but in some cases they can be more expensive. So, naturally, people ask the question: With all this money, do we really need this measurement?”
As a result, manufacturers know that existing metrics are inadequate, but it is not clear which new metrics justify the additional cost. Naravane says the uncertainty leads to wasted lab spending and contradictory expert advice.
Machine learning for measurement selection
Grainge AI's approach uses machine learning models to identify the most informative data points for a specific manufacturing problem. The company develops software tools that integrate with customer data infrastructure to answer measurement selection questions.
“Machine learning models allow us to learn which data is most useful,” Simmons says. “These provide a very natural solution to figuring out what data is relevant to this problem in a way that would take humans many times more time.”
The company's approach differs from recent food AI applications that rely on large language models like ChatGPT. LLM-based systems can automate some business processes, but Simmons points out that they face reliability challenges with failure rates of about 15%.
“When it comes to problem formulation, where you need the highest precision and reliability, you need solutions that are based on the data available from the problem you are working on,” he says. The company uses LLM agents internally for some tasks, but relies on data-driven machine learning for core manufacturing applications.
Focus on real-time manufacturing
The startup focuses on real-time production challenges rather than new product development. Naravane describes a scenario where a manufacturer receives a new batch of raw materials and needs to quickly adjust the formulation.
“We are solving manufacturing efficiency problems, which means in real time that there is very little time to solve problems,” she says. “It’s more of a real-time issue.”
The company targets mid-sized co-manufacturers and raw material suppliers facing formulation challenges. This includes maintaining consistent product performance and nutritional labels despite raw material variations, and determining the optimal use of non-traditional raw materials from side streams.
For data collection, Grainge AI integrates available customer data and works with a network of contract research organizations that can perform relevant measurements. The company also develops partnerships with method development laboratories that create high-quality testing protocols.
“They are not our customers, but rather our collaborators,” says Naravane. “We can think ahead about what these new use cases and new ingredients will be.”
Industry perception challenges
Beyond the typical startup obstacles to customer acquisition, the founders identify misconceptions about AI capabilities as a key challenge. Many companies dismiss AI applications because of limited exposure to language models.
“Perceptions of AI can be a little one-sided,” Simmons says. “People come to the table with an idea of what AI is, but that's probably 5% of what AI can actually do, or 5% of the forms AI can take. They may have erased in their minds the possibility that AI could work, when in fact another kind of AI may already be solving the problem.”
He points out that competition for the most human-like AI systems does not necessarily correlate with business success. “Depending on the business problem, you may not actually need to do that,” he says.
Origin and future direction
Mr. Naravane's interest is in managing food service operations in Germany. There, she was responsible for events ranging from 500 to 10,000 people and was faced with constantly reinventing recipes while minimizing food waste costs.
Rather than pursue a consulting career, she sought a data-driven technical approach and studied food chemistry at the University of California, Davis, USA. She met Simmons, who had been working on the subject of AI since 2018, at the university's AI Food Systems Laboratory.
The company is currently focusing on texture applications. It was chosen by Naravane because inappropriate textures can hinder product functionality on the production line. But she recognizes that flavor profiles are the frontier of the future.
Looking to the future, Naravane is focused on comprehensive ingredient understanding. “We have waste streams, we use materials to do things that have never been done before, and we find new sources of materials,” she says. “Understanding how ingredients work is a very deep subject and opens up many applications.”
Simmons aims to transform the compounding process from a reactive chore to a proactive tool. “I want the food industry to get to a place where food formulation is no longer a daunting chore that has to be done after the fact because the environment has changed,” he says. “It has to be easy to implement, proactively implement whenever you want, and be confident that the resulting solution is the right one.”
He cited near-term milestones as achieving this transformation within one to two years for several key customers.
Naravane added that data privacy remains a priority. “We are very responsible with the data we get from our customers,” she says. “We know that privacy and security are extremely important to all of our customers.”
