Nuclear's search problem gets a fleetwide answer trained on 53 million NRC pages

Nuclear stations across North America have started using NIVA, a retrieval system built by Atomic Canyon with the Electric Power Research Institute, the Institute of Nuclear Power Operations and the Nuclear Energy Institute ^[1][2]. It gives engineers and technicians across the commercial reactor fleet an automated way to query decades of technical, operational and regulatory records ^[3].

The detail that matters is upstream of the interface. The system runs on a model family called FERMI, trained with Oak Ridge National Laboratory on the Frontier exascale supercomputer, using more than 53 million pages of US Nuclear Regulatory Commission documentation ^[4][5][6]. According to Atomic Canyon, general-purpose language systems routinely fail on the dense terminology, precise abbreviations and domain syntax of nuclear engineering, and FERMI was built to interpret technical meaning rather than match keyword strings ^[7][8]. The completed model weights are published on Hugging Face ^[9].

That is the operative claim for anyone building retrieval in a regulated corpus: the vocabulary is the product. A licensing basis is not prose. It is a controlled language of NUREG numbers, procedure identifiers and abbreviations whose expansions are load-bearing, and a model that has never seen 53 million pages of it will guess ^[6][7].

The deployment is split in two. NIVA applies the retrieval models across broader sector archives, while a software environment called Neutron connects the same capability to an individual station's private network, letting local engineering run searches over its own licensing basis, maintenance logs, technical drawings and operating procedures ^[10][11].

Two modules are live. The Knowledge Assistant answers conversational queries from NRC Regulatory Guides, NUREG reports, NEI guidelines, INPO standards and EPRI research, with direct citations in every output so engineers can check statements against source documents ^[12][13]. The Operating Experience Assistant searches historical performance logs and prior event reports by context rather than surface word matches ^[14]. A third tool for diagnostics and troubleshooting is still in development, with plant trials planned for later this year ^[15]. One of the three announced tools therefore remains unreleased ^[16].

The commercial rollout follows a six-month pilot at several operating utilities, including Constellation Energy, which runs 21 reactors at 12 station sites ^[17][18]. Atomic Canyon had earlier tested generative tools on site at Diablo Canyon, operated by Pacific Gas and Electric ^[19]. An investment round backed by NVIDIA, Plug and Play Ventures and former Vanguard chairman Mortimer Buckley is funding expansion to additional sites, though no figure for the round was given ^[20].

Trey Lauderdale, Atomic Canyon's founder and CEO, said fleetwide availability shows AI in nuclear power is "real, operational, and ready to be deployed responsibly at a fleetwide scale" ^[21]. That is a statement about distribution, not about accuracy. No retrieval benchmarks, error rates or time-saved measurements appear in the announcement ^[22], and inline citation is a verification affordance rather than a correctness guarantee ^[13].

Three things to watch. Whether the diagnostics module clears plant trials this year, since troubleshooting advice carries different consequences than document lookup ^[15]. Whether performance holds on station-private material, given that the training corpus is regulator-side documentation while Neutron indexes utility maintenance logs and drawings ^[6][11]. And whether the published weights draw independent evaluation from anyone outside the founding consortium ^[9][2].