NukeML-in-a-Rack

A secure, modular, on-premise way to put NukeML on a nuclear site.

NukeML.ai

A deployment pattern, not a hardware lock-in

NukeML-in-a-Rack is a secure, on-premise way to adopt AI at a nuclear site. The platform runs locally. It can start as a reference rack, on approved customer hardware, or on qualifying infrastructure already at the site.

What it is

AI services that stay on site, with local compute and storage sized to the plant.

Why it matters

Nuclear sites need isolation, data sovereignty, auditability, and a clear IT/OT boundary before AI is scaled.

How it starts

The first year is operated on site. Approved data fills the knowledge base, and plant staff learn to run and maintain NukeML.

Four ways to deploy

The rack is the recommended reference model because it is the most controlled way to start. It is not the only way to deploy NukeML.

Reference rack

A preconfigured onsite system with compute, storage, staging, security, and operations patterns.

Approved hardware

The same platform on hardware acquired through the site’s own standards and suppliers.

Existing infrastructure

The NukeML stack on qualifying on-premise infrastructure, where cybersecurity and performance allow.

Phased footprint

Start with a smaller module set, then scale compute, storage, and services as data coverage grows.

Beside current plant systems

NukeML is designed to operate beside existing systems, not inside the control path. Plant systems such as SCADA, DCS, the historian, the work management system, and document repositories stay in place. Integration is an approved read-only export or pull. There is no write-back to operational technology, and new material is staged and validated before it is indexed.

The default posture is air-gap capable. Any outside connection is a site decision: approved endpoints only, encrypted, logged, and revocable by site administrators.

Build the knowledge base first

The platform becomes useful after approved data is captured, cleansed, indexed, and validated. Turning a module on follows data coverage, quality, traceability, access control, and a site-approved use case.

  1. Capture. Approved private site data, regulatory and public sources, drawings, procedures, and asset records.
  2. Cleanse. Normalize, deduplicate, chunk, enrich, and preserve the meaning of the source documents.
  3. Index. Build semantic, graph, and relational stores with configuration metadata and lineage.
  4. Validate. Quality gates, source citations, role checks, versioning, and owner signoff.
  5. Activate. Enable the available modules, then add later services when the data is ready.

Modules come on in phases

Initial modules

  • Nuclear AI Assistant
  • Intelligent Drawings
  • Intelligent Data Cleansing

Pilot or production use follows site validation and data readiness.

After validation

  • Asset Management + Digital Twin
  • Intelligent Training Platform

Activated after the model, workflow, data, and governance checks.

Later expansion

  • Further NukeML modules
  • Site workflows and agents

The same access control, audit, and source-traceability model stays in place.

Reference rack

The reference rack is a preconfigured onsite pattern. Sizing, processor count, storage, network boundaries, power, cooling, and the procurement route are set during the site assessment.

  • Perimeter. Firewall and network segmentation for controlled ingress and egress.
  • Compute. AI servers, staging nodes, and an application tier sized to the site.
  • Storage. Source documents, models, search stores, and audit logs.
  • Facility. Cooling and power matched to a sustained AI workload, using the site’s approved design.

Security

  • Air-gap capable. The default design supports no external connectivity and keeps inference local.
  • Separated zones. IT, operational technology, and the AI environment use least-privilege traffic rules.
  • Data stays on site. Private plant data remains onsite unless the site explicitly approves otherwise.
  • Audit. Queries, user actions, outside calls, and source access are logged.

The first year

The first year moves the platform from an onsite-operated start to site-enabled operations. Staff leave the year able to operate and maintain NukeML with written procedures, a data-refresh cadence, an incident model, and a support path.

  1. Months 0–3. Install and isolate, set up data access, begin ingestion, and establish the admin model.
  2. Months 4–8. Run data quality gates, pilot the first modules, onboard users, and set support routines.
  3. Months 9–12. Train operations and maintenance staff, finish admin playbooks, and move to steady support.

Pilot sequence

  1. Site readiness. Confirm infrastructure, security boundaries, data sources, stakeholders, and pilot scope.
  2. Deployment option. Choose the reference rack, approved hardware, or existing infrastructure.
  3. Data governance. Set the access model, approved sources, private-data handling, audit rules, and signoffs.
  4. First-year operating plan. Agree onsite staffing, onboarding, training, transition, and support.
  5. Module roadmap. Start with the Assistant, Drawings, and Data Cleansing. Add later modules after validation.

A pilot is successful when data coverage is validated, people are using it, answers stay traceable to sources, and the connection does not disturb current operations.