Sovereign Air-Gapped Engineering AI
Your engineers are watching the AI revolution through glass. The ban is right — the conclusion isn’t.
Every defence engineering organisation lives the same contradiction right now. On one side: the most knowledge-intensive engineering there is — decades of design history, works standards, configuration rules — and a workforce that is retiring faster than it can be replaced, while production ramps to levels not seen in a generation. On the other: a categorical ban on the tools every other industry now uses daily. ChatGPT, Copilot, cloud AI of any flavour — off limits, and rightly so. VS-NfD, export control, customer NDAs and plain common sense all point the same way: this data does not leave the building.
The wrong conclusion is that AI is therefore not for you. The right one is the one Europe’s defence sector itself is drawing — sovereign AI, exemplified by partnerships like Helsing and Mistral: if the cloud can’t come to the data, the model comes inside the perimeter.
That is what we build: engineering AI that runs entirely on your infrastructure — open-weight LLMs on your GPUs, retrieval over your Teamcenter and document bases, permission- and clearance-aware, with audit logs you own and zero outbound traffic. The productivity leap, without the data leaving home.
// Sovereignty, in numbers
The double bind: maximum knowledge intensity, maximum data restriction
Defence engineering is where the two curves cross at their steepest. Systems live thirty to fifty years, which means today’s engineers constantly need decisions, drawings and rationale from decades past. The people who carry that context are retiring in waves, exactly while the Zeitenwende ramp-up demands more throughput from design offices than they have delivered in decades. And the fastest competitors — the drone and NewDefence companies — were born with AI in their toolchain and no legacy constraints on using it.
Meanwhile your data rules are non-negotiable, and should be: VS-NfD material stays in accredited environments; ITAR- and AL-touched design data must never transit uncontrolled infrastructure; customer programmes carry NDAs with teeth. A cloud AI subscription is not a grey area here — it is a reportable incident waiting to happen. So the ban stands, the productivity gap grows, and your engineers privately wonder why the tools their industry needs most are the ones they alone cannot have.
Air-gapped AI dissolves the bind instead of picking a side: the models are open-weight and run on your hardware, retrieval happens against your indexes, access follows your clearance model, and the only thing that ever crosses the perimeter is an update you carry in — reviewed, on controlled media, on your schedule.
// The architecture, in one picture
Engineers
Ask in NX side panel, browser or terminal — SSO, clearance-aware access
Teamcenter · NX · shares
Sources indexed with permissions mirrored per user and marking
Open-weight LLM on your GPUs
Mistral- or Llama-class models, running on hardware you own — retrieval, generation and reasoning entirely local
RAG index
Your documents, embedded and searchable — never exported anywhere
Audit log
Every question, every source touched — logged for your security officer, owned by you
Public AI services, model vendors' telemetry, "anonymous" usage analytics — everything that makes cloud AI a non-starter for defence data.
Updates enter the other way: reviewed artifacts, carried in on controlled media, on your schedule.
Sovereignty is an architecture property, not a checkbox in a vendor's DPA. There is no outbound route to misconfigure, no telemetry to trust someone to disable, no "EU region" that is still someone else's computer.
What runs inside the perimeter
The air gap is the foundation — what earns its keep is what runs on it. Every engineering AI capability we build elsewhere is delivered here in its sovereign form:
Engineering knowledge assistant, air-gapped
Sourced answers from works standards, design reviews, calc reports and Teamcenter — the assistant your engineers see everywhere else, running where VS-NfD material may actually live. Onboarding accelerates, retiring knowledge stays.
Legacy drawings & obsolescence, mined locally
Decades of programme drawings — including reactivated production lines — extracted and indexed without a single scan leaving the building. The naval and land-systems archive problem, solved inside the fence.
NX automation with a local co-developer
NXOpen development and legacy macro rescue, LLM-accelerated — with the model on your GPUs, so even your automation code and the works logic it encodes never transit anyone's cloud.
Documentation & evidence drafting
Configuration records, change documentation, technical report drafts — generated locally from your PLM data, reviewed and released by your engineers. The paperwork burden of MIL-grade traceability, reduced at the source.
How a sovereign deployment works
Air-gapped AI is an infrastructure project as much as an AI project — and we treat it as one, together with your IT and your security organisation from day one:
Data zones & clearance model
- Which sources, which markings
- Who may ask what, see what
- Security org at the table early
- Hardware sizing from real load
The stack, inside the fence
- Open-weight LLMs (Mistral/Llama-class)
- Your GPUs — from one server up
- Vector index, RAG pipeline, UI
- No phone-home, verified
Sources, permission-mirrored
- Teamcenter, shares, standards
- Access per user, per marking
- SSO / your identity provider
- Audit logging from day one
Yours to run, ours to support
- Updates via controlled media
- Model upgrades on your schedule
- Quality measured, not assumed
- Your admins trained & enabled
Who this is for
You are a candidate if most of these apply:
- You handle VS-NfD, export-controlled (AL/ITAR) or NDA-restricted engineering data — and public AI tools are consequently banned.
- Design offices are under ramp-up pressure while experienced engineers retire — the knowledge-transfer problem has a deadline.
- You are a naval or land-systems supplier, a drone/NewDefence company, or a manufacturer entering the defence supply chain and meeting its data rules for the first time.
- Decades of programme documentation and drawings exist that only veterans can navigate.
- You run Siemens NX and Teamcenter (or comparable) and want AI integrated there — not another disconnected tool.
- Your security organisation demands architecture-level answers, not vendor assurances.
The sovereign stack
Every component chosen so that ownership, auditability and the air gap survive contact with reality.
Why inteliScience — and two honest sentences most vendors skip
Plenty of IT houses can rack a GPU server and start an open-source model on it. The hard part is everything after: making the assistant genuinely useful on engineering data — Teamcenter structures, works standards, drawing archives, NXOpen code — and making the access model track your clearance reality per user and per marking. That combination of Siemens-stack depth and AI engineering is our core business; the sovereign deployment is the delivery form your sector requires.
First honest sentence: an air-gapped open-weight model is not the largest frontier model. The gap has narrowed dramatically and keeps narrowing — but what actually decides usefulness is not raw model size, it is retrieval quality over your documents and integration into your tools. A works-specific assistant on a strong local model beats a frontier model you are not allowed to use, every single day. In the pilot we measure exactly this, on your questions, so the decision rests on evidence rather than vibes.
Second honest sentence: we don’t sell accreditation. Formal approval of systems in classified environments belongs to your security organisation and its processes. What we deliver is an architecture designed for that scrutiny — no outbound routes, full audit trail, documented components — plus the technical documentation pack your Geheimschutzbeauftragter needs to run the process. We make the approval defensible; your organisation makes it official.
Fixed price — no day rates, no surprises
Sovereign AI Pilot
- Full stack deployed on your hardware — model, RAG, UI, audit logging, verified zero-outbound
- 2 knowledge sources connected, permissions mirrored (e.g. works standards + one Teamcenter area)
- Evaluation on 100 real engineering questions — grounded-answer rate measured
- Security documentation pack for your Geheimschutz/IT-security process
- Rollout blueprint: more sources, more use cases, hardware scaling path
8–10 weeks · GPU hardware provided by you or quoted separately · everything deployed is yours, whatever you decide afterwards.
For comparison
The cost of the AI ban
- Engineers spending a quarter of their day searching what an assistant answers in seconds
- Programme knowledge retiring undocumented, one veteran at a time
- And the strategic one: NewDefence competitors iterating with AI while your ban stands still
Frequently asked questions
How good is a local model compared to ChatGPT — honestly?
How do updates work if nothing crosses the perimeter?
Can this be approved for VS-NfD environments?
What about ITAR- and AL-marked data inside the system?
What hardware does this actually need?
Siemens offers Teamcenter Copilot with an on-prem option. Why you?
The ban protects your data. It shouldn't cost you the decade.
Bring your security officer to the first call — we'd rather answer the hard architecture questions on day one than on day ninety.
Talk to a Sovereign AI EngineerAlways here to support you – reach out to us with your questions, and we’ll guide you every step of the way.
+40 774 463 429
+40 774 463 429
Romania
info@inteliscience.net
info@inteliscience.net
