From Bits to Steel

Building trusted AI for nuclear engineering and manufacturing at Everstar

Srijith Rajamohan, Ph.D., Chief AI Officer, Everstar

Introduction

Everstar's mission is energy abundance. Nuclear is the densest and most dispatchable source available to us, and it is badly underused relative to what it can do. It is also one of the most heavily regulated industries there is, and 10 CFR is where that shows up - the safety requirements are what make it trustworthy. They also make building anything new slow, understandably so, since every project carries a heavy load of analysis, documentation, and review. That load is where AI can help. Several of the topics below deserve more than a section each, and we will come back to them in later posts.

The demand picture sets our pace. Data centers alone went from 4.4% of US electricity in 2023 to a projected 6.7-12% by 2028, and electrification and reshored manufacturing are adding to that this decade. Meanwhile China and South Korea complete reactor builds in a median of under six years, against roughly a decade for Vogtle 3 and 4, the most recent US build. The difference is not physics or engineering knowledge; it is industrial process. That makes this an execution problem, which is the kind AI can help with.

Why nuclear is slow

Large reactors are power-dense and licensable under existing frameworks. Whether they beat fossil generation on cost is an execution question, not a research one.

Nuclear moves slowly, and the reason usually given for it is wrong. The industry is not slow because it is careless with rigor. It is slow because the rigor is executed by hand - by memory, by experts re-deriving the same judgments across thousands of pages, by review cycles that start over whenever a design changes. The rigor itself is not the problem; it is why the existing fleet has run safely for decades. The problem is that we have never had a way to execute it at anything other than human speed.

Decompose that and you get six bottlenecks that at first look like six unrelated disciplines: engineering and licensing, qualification of materials and parts, a workforce whose expertise is retiring faster than it can be transferred, manufacturing quality, construction, and operations that still run largely by hand. They share more than they appear to. All six are judgment-heavy, evidence-heavy, and require traceability. In each case the work means establishing a defensible position, sourcing every claim back to authoritative evidence, and being able to defend it years later to a regulator, an auditor, or an inspector. And in each case the evidence sits in decades of unstructured documents: scanned drawings, vendor submittals, procedures, calculation packages, and revision histories.

Those are the two halves of the problem. The bits are the filings, calculations, licensing bases, work packages, and change assessments that get produced and re-derived by hand. The steel is fabrication: welding, inspection, and nondestructive examination sit on the critical path, labor-intensive and qualification-bound, and that is where module throughput is actually lost. Two different regimes, both bottlenecked on rigor executed by hand.

Trust is the first constraint

In this domain, capability alone is not the bar. A principle I carried in from earlier work in another regulated field: a product that provides no response is better than a product that provides an incorrect response. In nuclear the cost of a wrong answer is a rejected filing, a schedule slip measured in quarters, or a weld that has to be cut back out and redone.

The harness is where that trust gets built

Choosing and adapting the right large language model (LLM) matters a great deal, but on its own an LLM is not a solution. A large share of our engineering effort goes into the agentic harness around it. Gordian, our platform for engineering, licensing, and operations work, is that harness: it decomposes a task, plans a route through it, calls the right engineering tools, checks its own intermediate results, and assembles the evidence trail behind whatever it concludes. Around the LLM sit expert functions, engineering and physics tooling, business rules, retrieval over curated corpora, verification chains, and a planner.

Building that harness for long-horizon work is the core capability of this team. A single task can mean working across hundreds of documents and dozens of calculations over several hours, where an error at step four quietly contaminates everything downstream. Agents at that horizon drift, and the drift is rarely loud. That is measured, not anecdotal: success rates fall from near-perfect on tasks that take a human minutes to under 10% on tasks that take hours, and models that commit to a wrong assumption early tend not to recover from it. Getting useful work out of them takes orchestration, tool interfaces, state management, checkpointing, intermediate verification, evaluation, and the discipline to fail closed rather than continue confidently. In a survey of multi-agent failures, task verification is one of the three main categories. This is most of what separates a demo from something a nuclear engineer will put their name on.

Four commitments the harness has to meet

Compute rather than recall. Where a calculation is called for, it gets calculated. The LLM contributes reasoning and judgment; it does not contribute arithmetic, and it is never the final authority on a technical claim. That is what the expert functions and engineering tooling in the harness are for.

Traceability by construction. Every assertion carries a path back to its source, and tracing pipelines record what the agent did and why. This is not a compliance layer added at the end. For work that has to survive regulatory review, the evidence chain is most of the deliverable.

Abstention as a capability. An agent that reports "I cannot establish this from the available evidence, and here is the gap" is doing what we want. Calibrated refusal is something we build and measure rather than something we tolerate.

Human review placed deliberately. Verification goes where the consequence is highest rather than everywhere, because review capacity is the scarcest resource in this industry. Spread it uniformly and you end up no faster than before.

This is also why we lean on first-principles work rather than pattern matching. You cannot pattern-match your way to a defensible engineering position, and you cannot benchmark your way to knowing whether you have one.

The same problem in metal

Titan, our manufacturing program, is where the same harness stops producing documents and starts driving machines. In fabrication it plans and monitors welding parameters in process, runs inspection continuously rather than as a gate at the end, and records provenance per part, from raw material to installed component. Quality gets built into the part rather than inspected into it afterward.

The commitments carry over with sharper consequences. Traceability becomes a per-part record read years later at component acceptance; abstention becomes halting a cell on inconclusive sensor evidence rather than laying down another pass. In software a wrong answer costs a retry; on a production line it can cost a component, a schedule, and even a certification.

Domain knowledge is where the value is

Frontier LLMs are very good, and getting better on a schedule none of us controls. What they cannot supply is knowing what correct means in this domain.

Concretely, domain knowledge is load-bearing in at least four places:

1. Data curation. The authoritative sources are specific and hierarchical: NRC regulations and guidance in 10 CFR, regulatory guides and NUREGs; the ASME Boiler and Pressure Vessel Code, Section III for nuclear components and Section IX for welding qualification; NQA-1 for quality assurance; IEEE and ASTM standards; EPRI and national laboratory research; and plant-specific documents such as safety analysis reports, technical specifications, and design control documents. Knowing which of these governs, which revision applies, and what supersedes what is a large share of the task. When you prioritize quality over quantity, curation ends up part art and part science.
2. Evaluation. There is no leaderboard for nuclear engineering judgment. We build our own evaluation toolkits, and grading them requires people who have done the work and can tell a defensible answer from a plausible-sounding one. Public benchmarks correlate weakly with downstream performance in a specialized domain, a finding that held in accounting too, and inspecting outputs closely, case by case, is eye-opening.
3. Output shape. The work has to land inside existing engineering and regulatory workflows, in a form the reviewer expects. Otherwise it can be correct and still go unused.
4. Where the human goes. Only domain expertise tells you which step is the one that must not be wrong.

Where commercial LLMs fall short in a specialized field, the problem is rarely fluency; it is the absence of a nuanced understanding of the domain. Nuclear leaves less margin for that than most. It is easy to believe that agentic AI is now an integration exercise, but almost nothing is plug-and-play. The value gets created in the last mile: the domain-specific scaffolding our engineers build around the LLM.

How we work

That last point is the practical reason our team looks the way it does. Nuclear engineers and licensing professionals who have carried real submissions and manufacturing programs work alongside the AI team during development, not as a review gate at the end. Domain judgment has to end up inside the system - in the agentic architecture, the eval design, the data decisions, the abstention thresholds, the placement of the human. And the only way to get it there is to build with the people who hold it.

What we are trying to do is decouple rigor from delay: keep the standard that earns this industry public trust, and stop paying for it in decades. The bits and the steel are the same problem approached from two ends, and closing the distance between them is the actual job.

Nuclear is hard and AI is hard, and neither is a reason to relax the standard. If these problems interest you, we are hiring.