AI Cuts Troubleshooting Time for FDA’s Food Safety Platform

As food safety systems continue to grow more complex, AI-assisted investigation paired with human oversight is helping to ensure the FDA’s GalaxyTrakr remains available when outbreaks demand the fastest possible response. 

The GalaxyTrakr platform sits at the center of that effort, allowing microbiologists to analyze DNA fingerprints from pathogens found in food samples and determine whether they are linked to outbreaks. The platform helps scientists identify clusters of bacteria, compare them against previous cases and quickly determine whether contaminated food should be pulled from the market. 

“We need to be able to make those decisions quickly. This data is very actionable,” said Errol Strain, senior information biologist at the FDA, at the AWS Summit in Washington, D.C. Tuesday. “If we identify a contaminated product that’s making somebody sick, we want to get it off the market as quickly as we can.” 

GalaxyTrakr is the FDA’s version of the public Galaxy bioinformatics analysis platform, customized with workflows designed specifically for foodborne pathogens. According to Strain, FDA laboratories process roughly 40 to 60 bacterial isolates each week while simultaneously tracking five to 10 active foodborne outbreaks.

“Galaxy lets us put wrappers and interfaces in containerized workflows, so that those people who are doing that analysis in the lab can just upload the data, push a button, go back to process some samples in the lab, and get a report out,” Strain said. 

But when those workflows fail, diagnosing the problem has historically been slow.

Because GalaxyTrakr relies heavily on open-source software and bioinformatics tools, failures can originate almost anywhere, such as improperly uploaded data, user mistakes, software bugs, infrastructure problems and storage limitations. Troubleshooting workflows that might involve 10 to 15 steps often require engineers to manually sift through logs spread across multiple AWS services and Galaxy components. 

“We can’t take days to solve that problem,” Strain said. “We really need hours to get the system back up to triage the data.” 

Jimmy Sanders, one of the lead architects for deploying GalaxyTrakr, said the platform has grown dramatically since its deployment in 2017. What began with fewer than 20 users now supports more than 500 active users who collectively submit about 80,000 jobs through an architecture spanning roughly 10 AWS services.

But that growth came with additional challenges, Sanders said.  

“Many times we’d have some active incident, and it would take days, if not almost a week sometimes, to really hone in on a particular issue,” he said.

The challenge wasn’t simply finding broken services, it was correlating logs, infrastructure, open-source code and documentation across an increasingly complex ecosystem. To address that problem, the team built an AI operations platform using retrieval-augmented generation (RAG) and the model context protocol (MCP). 

Jeremiah Jacquet, lead AI architect on the project, said the AI tool was able to retrieve trusted organizational knowledge including runbooks, documentation, source code and historical incident reports while providing standardized access to operational systems, monitoring tools and databases.  

“Responses are grounded in real organizational knowledge and can include citations, which increase accuracy and trust,” he said. 

Jacquet emphasized, however, the importance of human oversight in the process and gave an example of the AI identifying a GalaxyTrakr configuration issue and software defect. But the system stopped short of making changes on its own. He added that the goal is not autonomous decision-making, but rather accelerated decision-making. 

“The AI did not automatically deploy a fix,” Jacquet said. “The proposed solution was reviewed by engineers. The code was examined. The remediation was tested. Validation occurred before deployment.”