Meet NIVA: The nuclear industry’s own AI tool

Built on decades of technical, regulatory, and operational experience, the Nuclear Industry Virtual Assistant (NIVA) is out of beta testing and in use at nuclear power plants across North America.

NIVA debuted in August, featuring plain-language generative AI functions that would be familiar to people who have come to rely on AI chat tools. But NIVA's tools, including a "knowledge assistant" and an "operating experience assistant," are industry-specific and designed to help utilities accelerate licensing and maintenance. NIVA can make both historical lessons learned and the latest regulatory guides more accessible. It can also help newcomers to the nuclear workforce quickly learn what they need to know.

What do people in the nuclear industry need to know about NIVA? It begins with how NIVA was conceived as a collaborative effort by three organizations: the Electric Power Research Institute, the Institute of Nuclear Power Operations, and the Nuclear Energy Institute. Nuclear News went to project leaders to learn more.

Of Nuclear, For Nuclear

Leading the project for their respective organizations were Robert Austin, a senior program manager for operating plant initiatives at EPRI; Ramon Martinez, vice president of data and analytics at INPO; and Kelli Voelsing, senior director of engineering and risk at NEI.

"ChatGPT came out in late November 2022," Voelsing said. "And very quickly, our organizations, as well as all of our member organizations, started to think, 'How can I use this?'"

INPO, EPRI, and NEI each have troves of documents. Collectively, those databases are "a gold mine," Austin said, and people working in utilities have been expected to retrieve information from each organization's separate databases. But the task of searching for necessary information was laborious. "It's multiple databases with multiple search techniques, and sometimes stuff is hard to find," Austin said.

INPO's Martinez explained that during an industry meeting, Constellation CEO Joseph Dominguez challenged INPO, EPRI, and NEI to seize the opportunity to build a united approach to AI, connecting data sources to benefit the whole industry. Dominguez decided to fund an initial pilot of an industry effort co-led by the three organizations. "Not an INPO product, not an NEI product or EPRI product," Martinez said.

"Everyone was very hesitant about, or in some cases just unwilling to, put industry data that is closely held into just any AI tool, either for reasons of confidentiality or concerns about export control or for proprietary reasons," Austin said. Despite-or perhaps because of-that concern, they decided to work together. "The driving force was, what if we had an industry-specific AI system that would basically be a safe haven for all this information?"

With the chief nuclear officers of INPO, EPRI, and NEI-respectively, Jeff Place, Steve Swilley, and Doug True-fully on board, the organizations set out to produce a tool for an entire industry. While INPO sponsored the project, EPRI provided project management, and NIVA was created with an oversight committee consisting of INPO, EPRI, NEI, and Constellation (later, Pacific Gas & Electric, Southern, Ontario Power Generation, and NextEra Energy were also invited to join the committee).

Building the Framework

The project started in December 2024, but "the starting gun was really fired in January 2025," Austin said. At first, NIVA was simply known as "Nuclear Industry Large Language Model." One of the committee's first tasks was defining their goal and what they needed from a technology partner.

"We specified very clearly up front that we were uninterested in taking all this industry data, giving it to a language model, and 'fine tuning' it because of the concern that you might lose control over the data," Austin said. Instead, they chose to develop a retrieval augmented generation (RAG) tool that could use an LLM to query documents housed in databases and then feed query results through the LLM. After a competitive bidding process, Atomic Canyon was selected as a partner.

NIVA is built on multiple separate databases, and as such, its creators consider it a federated tool. "We knew that we probably would not be able to get all of the incoming reports and data onto a single database in a single location, both because of the risks but also logistically," Austin explained. "EPRI alone has over 7,000 reports."

"We wanted the ability to do this retrieval augmented generation-or RAG retrieval-over the internet in a way where the main NIVA framework is running at INPO but is accessing databases in different locations for different reports," he added.

NIVA users-that is, members of INPO, EPRI, and NEI-have access to the same information they already can access through the internet. If someone is a member of INPO but not of NEI, for example, they can't access NEI documents through NIVA. Martinez said, "If you didn't have access to that information beforehand, you're not going to have access to that information now, unless there is a change to the underlying agreements."

NIVA uses Atomic Canyon's FERMI AI models, which were designed specifically for search and retrieval of nuclear industry information. The models were trained using Oak Ridge National Laboratory's Frontier exascale supercomputer on more than 53 million pages of public Nuclear Regulatory Commission data. NIVA benefits from that nuclear-specific training while separately providing users access to a large collection of the NRC's publicly available regulatory guidance documents.

Atomic Canyon previously worked with Pacific Gas & Electric to deploy a generative AI system at Diablo Canyon. When NIVA was released in August, Atomic Canyon founder and CEO Trey Lauderdale noted that it was "a defining moment for the nuclear industry. The U.S. nuclear sector sits on decades of invaluable technical and operational knowledge, but too much of that knowledge remains difficult to access at the speed modern deployments require. The fleetwide availability of NIVA shows that AI in nuclear power is real, operational, and ready to be deployed responsibly at a fleetwide scale."

NIVA may later include data sets from industry support organizations such as the owners groups for pressurized water reactors and boiling water reactors. As Martinez explained, those groups "want to have assurances that they will maintain control of their data and will be able to interact and only provide access to those who are members of that organization, maintaining their contractual agreements and the intellectual property controls that they have in place."

One-Stop Shop

The choice of RAG architecture means NIVA's database is not static. "When EPRI releases a new report, INPO gets new OE [operating experience], or NEI publishes something, it gets added to that database and becomes searchable going forward," said Voelsing, explaining that search results are tailored to each plant.

That is helpful, especially because of frequent updates to NRC rules and guidance. Austin said, "Something NEI has worked a lot on with NIVA is making sure it's giving you the most up-to-date revision of a regulatory guide, for example, but also making sure you still have access to the older version that perhaps your plant is committed to in the license. This is key, and it's a subtlety that's been a little bit hard to do in the past."

Transparency and citations were an important design feature. "When it gives you an answer, it's going to give you the citation, and you're going to be able to click on that and download the source," Voelsing said. Summing up the value of NIVA, she added, "From an end-user perspective, instead of logging into five different systems to find out what documents are related and download them and upload them into my own system, now it's a one-stop shop."

But NIVA takes accessibility one step further. Early on, the collaborators knew they wanted an API (application programming interface) in addition to the traditional chat and search pages. Austin explained that NIVA supports this integration through Model Context Protocol, an open standard for connecting AI applications to external data sources and tools. NIVA's MCP allows utilities that are already using enterprise AI software to "hook NIVA as another tool or data source into their utility's Copilot, Gemini, Claude, or similar tool, making this more of a one-stop shop."

INPO collects confidential station-specific performance information through the Industry Reporting and Information System (IRIS), which is not held in the databases on which NIVA draws, Martinez explained. But by using the NIVA MCP-which in late September was being tested at EPRI-utilities can access their local, station-specific information and NIVA's broad resources in the same application.

INPO, EPRI, and NEI have briefed the NRC on NIVA. "They are developing their own internal AI system and trying to figure out how to use that to accelerate and make things more efficient," Voelsing said. "They're interested in understanding our lessons learned."

New Workforce, New Technologies

New companies are entering the nuclear industry, and both newcomers and established operators are planning new projects.

For companies that are seeing "a huge exponential growth in the number of workers needed at all levels of the organization and all different disciplines," NIVA is a knowledge transfer tool, Voelsing said. "When you come into a nuclear plant, you start hearing all these new acronyms and new terms and new processes. This offers a way to enter that space much more quickly. It provides a great orientation tool and a way to find the relevant references for the topics that you're coming across in the course of doing your job. We see strong interest in that across the operating fleet and across the advanced reactor and supplier landscape."

As new companies work on designing nuclear plants-even plants with significant differences from the light water reactors operating in the United States today-they can incorporate lessons learned. "I know a lot of their focus is on the primary system, the things that connect to the reactor," Martinez said. "But the secondary system is equally important."

From the early days of the nuclear industry in the 1960s, "Reliability has continued to improve in no small part because of the effort the industry has put into having high reliability and learning from past mistakes and eliminating single point vulnerabilities, eliminating challenges," Martinez said. "The industry has very high capacity factors, and I'm sure the new entrants want to achieve that as quickly as possible, not going through the same kind of growing pains that the traditional industry went through in the '70s and '80s."

But how can NIVA, with data sets built on LWR operations, help new reactor developers? Austin explained, "Even exotic reactors like molten salt still have a lot of common equipment that we have a lot of operating experience on." New companies "still have a balance of plant. . . . Some of them still have steam turbines, definitely have pumps, air compressors, valves, circuit breakers, electrical gear." For such common equipment, "even if the reactor's different, there's still hundreds or thousands of years of operating experience of this equipment in a nuclear utility environment available that's now accessible through NIVA," he said.

To Voelsing, "One of the best features about the operating experience tool is, as Ramon [Martinez] said, you type in what your issue is, and it comes back with a list, and the tool even categorizes and tells you, 'This is why I prioritize this high, this was a similar event, similar type of equipment, but it was a different type of reactor.' Or it might say 'this wasn't the same equipment, but it involved management or leadership.' And so it tells you why it picked certain things, which I think is extremely helpful."

Testing It Out

By the time a beta-testing pilot program began in early 2026, utilities were eager to give NIVA a try.

The pilot program included six utilities. "Constellation started, but then we added a few more utilities to the steering committee, really looking for a mix of utilities," Austin said. "We wanted to have a representative from a regulated fleet. We wanted to have another fleet that does a lot more of their own AI work. Everyone responded with subject matter experts and testing."

Constellation, with its fleet of 21 operational nuclear reactors, was an important player in beta testing. Brian Carroll, who leads the utility's nuclear digital strategy, worked closely with Austin, Martinez, Voelsing, and others to make NIVA a reality.

Of the collaboration and NIVA, Carroll said, "It reflects real challenges our workforce faces every day, and we're proud to be the first to help lead its adoption across the nuclear fleet."

NIVA cannot replace engineering judgment or perform calculations. Its creators caution that like all AI models, it can occasionally produce responses that sound confident but contain errors.

Also like other AI models, NIVA "expedites and accelerates," Austin said. "I've learned it's really good at taking OE reports that are written by engineers, or maybe event reports for the NRC, which are sometimes not written in the clearest language possible, and translating it into something that's much more understandable and therefore more accessible and therefore much more likely to be used."

One of Voelsing's favorite NIVA metrics came from a pilot program exit survey. Participants were asked if they would be disappointed if they no longer had access to NIVA. "The response was overwhelming," she said-82 percent of respondents said "they would be somewhat or very disappointed."

Another important takeaway from beta testing was how interested users were in adapting the tool for specific needs. Voelsing said they received a number of queries about adding new data sources and use cases, among other things. "The potential is that the system will continue to grow, and we've scoped it in a way that we can continue to grow. We can bring in more data sets and use cases, adapting it for different populations," she said.

NIVA was built with all INPO and World Association of Nuclear Operators Atlanta Center members in mind as its end users. That includes operating nuclear plants in Canada, Mexico, Romania, and South Africa. In late September, the United Arab Emirates's Barakah plant, also part of WANO Atlanta Center, was in the process of finalizing legal agreements to gain access.

Staying On Top of Changing Technology

NIVA's roll-out, which began in mid-August, has been going "spectacularly well," Austin said in September, with usage trending upward.

NIVA went through security testing by an independent company before it was declared production ready. It is deployed on a plant's business network, well isolated from plant protection systems.

"We are not running a frontier model," Austin said, but rather a standard AI model available on Azure-one that's not designed for agentic behavior like that displayed by experimental models with agents that went rogue and broke out of testing environments this summer. "The way NIVA is designed, it's not one of those models where we have to worry about it crawling out and crawling through the plant."

Soon, NIVA will feature a troubleshooting assistant that builds on the operational experience and knowledge transfer tools. In late September, the troubleshooting tool was in beta testing, with expectations that it would be rolled out before the end of the year.

"It's a more specific application and because it needed operating experience and knowledge transfer working, it's lagging a little bit behind the others, and we set that expectation up front," Austin said.

Through the troubleshooting tool, a user will be able to ask a specific question about plant operations, such as, "My pump motor just stopped and there's smoke coming out of the bearings, what could be causing this?" Austin explained.

NIVA's response would be modeled after typical industry support-refute tables and would list possibilities from "you forgot to put grease in the bearing to some sort of electrical issue, and then provide the 'refute'-the steps you need to take to determine whether or not this was the cause," Austin said. Solving the problem still requires a careful process of elimination, but "just assembling that support-refute table, which could have 50 or 100 rows of data, takes time. With the troubleshooting assistant, the intention is that we'll do that very quickly, within seconds or minutes, saving a lot of work."

This is the kind of task NIVA was built for. Carroll summed it up: "NIVA was built by the industry, for the industry."

Susan Gallier is the editor- in-chief of Nuclear News.