GE Aerospace Created AI Wingmate in Six Weeks. Can You Say ‘Next Level’?

In 2022, not long after ChatGPT was first made available to the general public, GE Aerospace leadership noticed many employees using the powerful generative artificial intelligence platform in their work. While this was unsurprising for the company’s tech-forward culture, it also raised concerns that sensitive or proprietary information might find its way into the data oceans such large language models (LLMs) train on. The leadership team saw two options. “One, we could simply block access to ChatGPT on work laptops,” recalls Dinakar Deshmukh, vice president of data science. “Or two, we could provide a safe, secure, internal alternative to ChatGPT that allowed them to continue using the tools we know they want and could benefit them and the company.” 

Clearly, option two was more attractive. The problem was that someone would have to build this GE Aerospace–only version of generative AI, make it as powerful, responsive, and resourceful as the consumer version, and do it quickly enough not to frustrate employees who had incorporated ChatGPT into their work life. This past spring, a small team partnered with Microsoft to attempt this very thing. 

“We put together a road map for how we were going to build it, as well as how we’d release it when it was ready,” says Deshmukh. Working with Microsoft’s accelerator program for LLM chat apps, they helped Microsoft adapt its codebase to meet GE Aerospace’s specific needs, making the company’s deep wellspring of knowledge instantly accessible to all employees through targeted searches in its chat function. 

Microsoft estimated that it would take two weeks to set up the application’s back-end service. When he heard this, Deshmukh chose a goal of four to six weeks for his GE Aerospace team to execute the rest. “Sometimes ignorance is bliss,” he says.

How the GE Aerospace Team Transformed Teamwork

The project to build this new internal platform — which they named AI Wingmate — drew on a vast array of cross-functional talent within the company. “A product manager, a data scientist, a user experience expert, cloud and back-end specialists, cyber, legal, middleware — people from all different parts of the organization were involved,” Deshmukh continues, citing active participants in Bengaluru, Cincinnati, Detroit, Florida, New York, and Seattle. “We’d have a morning conversation at nine o’clock and an evening one at five, because a team in India was working on the same issues. We deployed the ‘follow the sun model’ very effectively for this project. The first four weeks were extremely intense. Everybody was working well beyond their normal hours, because they all realized that giving something like this to GE Aerospace employees, at this scale, would be game-changing.”

Why game-changing? “Until now, there was no application available to our entire employee base,” he explains. “And by its nature, this will shift the culture within GE Aerospace, giving everyone the ability to interact with the AI systems that can take them to the next level.” For more than a decade, the company had been marshaling AI technology to enhance engine monitoring, predictive maintenance, fuel optimization, inventory optimization, and a host of other specialized functions, as well as partnering with other hyperscaler companies for specific business applications. But this would transform the work life of every employee at GE Aerospace. “Everyone on the team realized they could be a catalyst for this,” says Deshmukh. 

They still needed to overcome a range of challenges to execute the project at scale. “For one thing, we had to make sure that we developed this in a very compliant way, from both a cyber and legal standpoint,“ Deshmukh says. Since AI Wingmate operates within the guardrails of GE Aerospace’s virtual private cloud (VPC) network, every pipeline in and out of that VPC is encrypted. “We had to get approvals and follow that process in very little time without taking shortcuts.”

They also had to develop a unique GE Aerospace identity for the application. “Once we had the initial layouts, our branding team played a key role in helping us with finalizing the look and feel of the platform and making it real,” he continues. “The branding and communications team was instrumental in helping us come up with the name for the internal application, AI Wingmate.”

Learning How to Use GE Aerospace’s AI Wingmate

Once Microsoft built the back end, the GE Aerospace team created the system’s interface. “At first we had trouble making it work smoothly, because some responses from the GPT-4 model were unexpected,” says Deshmukh.

While fixing these issues, the team found a problem with the application’s responsiveness. “We saw that sharing large public graphic processing units (GPUs) with users worldwide was slowing the app down,” he says. To address this issue, Microsoft agreed to set aside some GPU capacity just for AI Wingmate. “That action on the part of Microsoft really helped resolve the problem, and from there we did not look back.”

From that point, AI Wingmate underwent a rapid, internally crowdsourced evolution, much like what machine learning itself does in the wild. “At four weeks, we released it to the first batch of 2,000 users and started getting feedback on what was working and what wasn’t working so well,” Deshmukh says. “We learned a ton and were able to make some important final adjustments to ready the tool for launch to all 52,000 GE Aerospace employees.” 

Rather than the free version of ChatGPT most employees were familiar with, the team decided to base AI Wingmate on the higher-end technology in GPT-4. “It’s a much higher order of cost, but we decided to give our employees the best possible experience,” says Deshmukh. They also ran and recorded training sessions to determine how best to interact with AI Wingmate, and made these videos available as guides for effective usage. 

One important aspect of generative AI is that “no matter what model you’re using, the responses you get are dependent on two important factors,” Deshmukh says. “One: how you ask the question. And two: how you validate the answers you get back. As people get more skilled at prompts, they’re able to use AI Wingmate to help them explain increasingly complex things in an easy-to-understand way.” A good teacher as well as communicator, AI Wingmate also helped some employees learn programming fundamentals. “I know many engineers are using it to learn Python,” says Deshmukh.

Failing, Learning, Adjusting, Then Moving Forward with FLIGHT DECK and More

Deshmukh reports that the system has topped more than 1 million prompts from 14,000 active users within the past 150-plus days. For the past five months, AI Wingmate has been available to all employees. The next phase will integrate FLIGHT DECK, GE Aerospace’s proprietary lean operating model, into the platform, providing a dedicated learning module to help employees engage with this powerful tool.

With the AI Wingmate project, GE Aerospace stands out from industry peers as an early adopter. “GE Aerospace is an industrial company,” says Deshmukh. “We are not a digital company. But at the same time, we want to make sure our workforce has a best-in-class experience, and we can provide them with all the right tools and resources for them to be most effective.” 

The rapid rollout reflects digital tech’s ever-evolving nature through learning and reiterations, as well as the sense of community within GE Aerospace. “Going to 52,000 employees with an AI application like this in six weeks might sound a bit crazy to some companies,” Deshmukh says. “But on this project, we were willing to fail, learn, and adjust ourselves rather than wait for eight months or a year to find the absolutely perfect solution for everyone.”