Almost a decade ago, while others experimenting with AI focused on algorithms for trading, diagnostics, or digital advertising, a company called NotCo was experimenting with AI by the name of Giuseppe to create plant-based foods that could match the taste and texture of their animal-based counterparts.
According to Aadit Patel, SVP of AI Product and Engineering at NotCo, the company’s founders (Patel would join a couple of years after the company was founded in 2015) realized early on that, in order to build an AI model that could help create plant-based products mimicking the taste, texture, and functionality of their animal-based counterparts, they would need a whole lot of data.
The problem was, as a startup, they didn’t have any.
When I asked Patel in a recent interview how the company overcame the infamous “cold start” problem—the challenge many embryonic AI models face before they have built large datasets on which to train—he told me they found the solution in a very public place: the U.S. government’s website.
“In the early days, when we had no money, we literally scraped the USDA website,” said Patel. “If you go to the USDA website, there’s a bunch of free data materials for you to use. And I guess no one had actually joined it together to create a comprehensive dataset… So the first versions of Giuseppe were built on that.”
This cobbled-together dataset formed the foundation for Giuseppe’s recommendations, leading to the creation of products like NotMilk, which uses unexpected combinations like pineapple and cabbage to replicate the taste of dairy milk.
As NotCo grew, so did Giuseppe’s capabilities. New analytical labs in San Francisco and Santiago, Chile, gave the company a wealth of new data on which to train its AI. Over time, the model’s ability to create innovative food products also improved.
One of the biggest hurdles in food development is the fragmented nature of the supply chain. Data is scattered across various entities—ingredient suppliers, flavor houses, manufacturers, and research institutions—each holding critical information that contributes to the success of a product. Over time, the company realized that to create an AI capable of building innovative products, it couldn’t rely solely on NotCo’s datasets. Instead, Giuseppe would need to integrate and analyze data from across this complex web of partners.
“What we’ve done with Giuseppe is figure out a way to incentivize this very fragmented ecosystem,” Patel said.
According to Patel, pulling together these disparate datasets from across the product development and supply chain would result in a more holistic understanding of what is needed for a successful product that is better aligned with market realities.
“We realized that if we just made an AI system that’s specific to CPG, we’d be losing out,” said Patel.
Generative AI and Flavor and Fragrance Development
One recent expansion of Giuseppe’s capabilities has been the exploration of new flavors and fragrances using generative AI. While GenAI models like ChatGPT have become infamous for creating sometimes strange and off-putting combinations when designing recipes and new food product formulations, Patel explained that the company has been able to overcome issues with general LLMs by creating what he calls a discernment layer. This layer filters and evaluates the multitude of generated possibilities, narrowing them down to the most promising candidates.
“Discernment is key because it’s not just about generating ideas; it’s about identifying the ones that are likely to succeed in the real world,” Patel said. “With generative AI, you can prompt it however you want and get an infinite amount of answers. The question is, how do we discern which of these 10,000 ideas are the ones most likely to work in a lab setting, a pilot setting, or beyond?”
The discernment layer works by incorporating additional data points and contextual knowledge into the model. For instance, it might consider a formulation’s scalability, cost-effectiveness, or alignment with consumer preferences. This layer also allows human experts to provide feedback and fine-tune the AI’s outputs, creating a process that combines AI’s creativity with the expertise of flavor and fragrance professionals.
Early tests have shown positive results. When tasked with creating a new flavor, both the AI and the human perfumers receive the same brief. When the results are compared in A/B tests, Patel says the outputs of Giuseppe’s generative AI were indistinguishable from those created by human experts.
“What we’ve built is a system where AI and human expertise complement each other,” said Patel. “This gives us the flexibility to create products that are not just theoretically possible but also market-ready.”
CPG Brands Still Have a Long Way to Go With AI-Enhanced Food Creation
Nearly a decade after building an AI model with scraped data from the USDA website, NotCo has evolved its AI to create new products through a collaborative approach that results in a modern generative AI model incorporating inputs from its partners up and down the food value chain. This collaborative approach is being used for internal product development and third-party CPG partners, many of whom Patel said approached the company after they announced their joint venture with Kraft Heinz.
“Ever since our announcement with Kraft Heinz and signing a joint venture, there’s been a lot of inbound interest from a lot of other large CPGs asking ‘What can you do for us?’ and ‘What is Giuseppe?’ They want to see it.”
When I told Patel I thought that big CPG brands have come a long way over the past twelve months in their embrace and planning for using AI, he slightly disagreed. He said that while there’s a lot of interest, most big brands haven’t actually transformed their business to fully create products with the help of AI.
“I would say there’s strong intent to adopt it, but I think there hasn’t been put forth like a concrete action plan to actually develop the first AI-enabled R&D workforce,” said Patel. “There is room, I think, for new AI tech for formulators, and room for best practices and lessons learned of adopting AI.”
You can watch my full interview with Aadit below.
The NotCo team will be at the Food AI Summit talking about their new efforts using generative AI to develop flavor and fragrance, so make sure to get your tickets here.
Leave a Reply