Bayer Crop Science blends gen AI and data science for innovative edge

Bayer Crop Science sees generative AI as a key catalyst for enabling thousands of its data scientists and engineers to innovate agricultural solutions for farmers across the globe.

The agricultural division of life sciences multinational Bayer is developing a new data science platform based on Amazon SageMaker Studio, says Will McQueen, head of Crop Science global data assets at Bayer. Infused with the gen AI capabilities of Amazon Bedrock and Amazon Q, the platform has been designed to facilitate and expedite the creation of “novel” agricultural products, he says.

A closeknit team of about 10 engineers and executives from Bayer, Amazon, and Slalom Consulting cooked up the blueprint for the “Decision Science Ecosystem” roughly 18 months ago and has been building the platform for about a year. Data scientists at Bayer have developed several proofs of concept of generative AI models on the new platform that remain in discovery and evaluation phase for “efficacy,” McQueen says, adding that the models won’t be in production until 2025.

“The R&D pipeline is pretty highly confidential at this point,” he says. But suffice to say, advanced generative AI could potentially one day lead to the creation of hybrid seeds or new seeds that could alter and add to the food supply chain. “The core of our company’s mission is feeding the world,” McQueen says.

Bayer Crop Science’s existing data science platform, built on a licensed platform dubbed Domino about 7 years ago, is running out of steam and needs to be replaced for the modern AI era, McQueen says. Plans for the first major release of Decision Science Ecosystem are within the next couple of months.

Like most enterprises, Bayer’s agricultural division will initially use AWS-based generative AI tools out-of-the-box to automate basic business processes, such as the production of internal technical documentation, McQueen says. The core set of engineers building the platform have harnessed this feature to speed up the process. Making that available across the division will spur more robust experimentation and innovation, he notes.

“Prior to this capability, individual engineers would have to create their own documentation from the code they’re writing and other development. AWS’ out-of-the-box capabilities replace this manual work, making our engineering staff more effective and able to deliver value a lot more quickly than before,” McQueen says. 

The forthcoming data science platform — which will be used by engineers and data scientists across Bayer — also features enhanced connections and integrations to Amazon Bedrock, the ability to write code using natural language, and robust testing and safety guardrails.  

The Crop Science team has also developed unique features on top of the data science platform, including its model registry, which is a custom catalog of AI models and model lifecycle features that track a model from discovery to testing, deployment, and production with requirements at each stage. The model registry also enables data scientists to leverage code developed by colleagues, McQueen says.

“From an acceleration perspective, it’s the reuse dimension that we see for creating opportunities using generative AI and Bedrock,” he says. “As the model is moving through its lifecycle, it has different stage gates and requirements that enforce those stages.”

Initiating change

The agricultural division has taken its generative AI efforts and partnership with AWS to a new level by developing this technical platform, which can not only facilitate model development but also train data scientists and engineers on prompt engineering techniques and the application of advanced data technologies to create new commercial offerings.

At its recent summit, AWS listed Bayer Crop Science among its enterprise customers pushing the envelope on generative AI innovation, alongside Exscientia’s advanced drug discovery platform and EvolutionaryScale’s ESM3, which gives scientist and biologists a prompt engineering platform for building various proteins for experimentation.  

Such innovation is built into the DNA of Bayer Crop Science’s scientists and engineers, but the level of change management involved in using gen AI platforms — even for the highly skilled — is complex and requires “thoughtful evaluation,” McQueen emphasizes.

“Being open to doing jobs fundamentally differently and leveraging AI takes a little bit of time to get used to,” he says.

Gen AI’s basic features — such as document summarization and content creation — are already enhancing the quality of the data science platform and reducing time to availability, McQueen says. But over time, Bayer’s innovators and agricultural scientists will become better adept at incorporating the platform’s unique tools and capabilities to innovate in unprecedented ways.

“This will help onboard employees onto [our] gen AI platform much more quickly and better understand its capabilities to build models,” McQueen says. “It’s additive.”

Still, doing so will require great oversight and robust quality control procedures, he says, acknowledging the risks that come with experimenting with the most advanced scientific tools on the planet. 

To that end, the Crop Division team has built in safeguards to prevent proprietary data from trickling out of the platform — or worse, deploying promising but untested solutions to Bayer’s global farming population. 

An open approach

Bayer Crop Science operates a multicloud environment, but McQueen opted to partner more closely with AWS on gen AI due to its more flexible, open platform, he says. The Bedrock-based platform will allow Bayer’s data scientists and engineers to access a variety of open-source large language models (LLMs) available from marketplaces such as Hugging Face.

Amazon’s AI platform also enables customers like Bayer to use their data platform of choice, a critical aspect of developing gen AI models. In Bayer Crop Science’s case, Google BigQuery is the division’s data warehouse, McQueen points out.

“One thing we identified with AWS early on was the ability to build technical capabilities that were flexible so that we could develop a more componentized architecture that allows us the ability to plug in different models from different providers,” he says.

Bedrock’s catalog of gen AI models, for instance, includes open source and closed models from partners, such as Meta’s Llama 2.1 and Mistral’s Large 2, its most advanced gen AI model.

Dave McCarthy, research vice president of cloud and edge services for worldwide infrastructure research at IDC, points out that cloud providers’ tacks for generative AI remain distinct.

“The cloud providers are going to take different approaches to making these different types of models accessible to their customers. Google has been focusing on the value of its internally developed Gemini family of models, whereas AWS has taken a more partner-led approach with third-party model vendors,” he says. “There is no clear winner as to which approach will be best in the long run.”

In the meantime, as enterprises move toward more advanced development of gen AI models, CIOs will have a lot to manage in terms of vendor partnerships, procurement, costs, development, measuring outcomes, and security.

Safeguarding the process

Although still early in its journey, Bayer Crop Science is actively developing a number of new use cases that McQueen claims will be “disruptive” in the agriculture industry.

The division is advancing its data science platform in parallel with Bedrock and plans additional launches that correspond with Amazon’s AI platform updates.

As they lean into advanced use cases, McQueen’s IT group has integrated automated filtering and monitoring tools and other safeguards to secure proprietary data, including integrating methodologies that guide scientists and engineers on responsible development.  

“With any new capabilities we’re developing that have the potential to go to market or directly be embedded in a workflow process, employees have to do careful benchmarking and testing of that capability before it is launched mainstream,” McQueen says of his team’s efforts to keep prototypes from sliding into the food supply without rigorous quality control testing built into the data science platform.

“We can test [models] side by side with human experts to do the validation before it reaches full production,” he says. “It’s important we don’t launch a new capability externally, and then give bad advice to a farmer that impacts their business and causes harm.”