Sovereign AI for industrial teams

Your technical documents.Verifiable AI answers.

We build private RAG systems for engineering and industry. Specifications, manuals, and quality records become securely searchable — with exact sources, existing roles, and audit logs.

On-premises or private cloudA source required for every answerIntegration with existing roles

Grounding Evaluation

RFQ-2026-0147 · Technical review

Private Boundary

Specification_Plant_A17.pdf

p. 42 / 186

Retrieved evidence

Mandatory criterion 4.2: The system must support continuous operation for at least 8,000 hours at 400 V ±10%.

Review question

Which mandatory operating-voltage requirement is not yet evidenced in our proposal?

Source-grounded answer

The current proposal does not evidence 400 V ±10% and at least 8,000 hours of continuous operation.

Source 01 · page 42 · § 4.2Grounding: passed
Document, retrieval, and model remain inside the defined security boundary

Synthetic evaluation case · no customer data

Deployment & control

On-premises ready
Private-cloud capable
Air-gapped capable
Sources & RBAC
The real risk

The biggest AI risk is not the model. It is uncontrolled data flow.

For specifications, RFQs, quality records, or CAD metadata, a plausible answer is not enough. Who processes the data, where it is stored, and whether every statement can later be verified are what matter.

RISK / 01

NDAs & intellectual property

Engineering knowledge and customer specifications remain inside the agreed infrastructure and processing boundaries.

RISK / 02

Controllable vendor risk

Models, hosting, telemetry, and data flows are selected deliberately, documented, and technically constrained.

RISK / 03

Evidenced, not merely plausible

Answers are grounded in approved sources and returned with exact references for human review.

Sovereign-by-design

From document to evidenced answer — under your control.

We do not blindly deploy a standard stack. Model, retrieval, hosting, and access controls follow the data class, existing IT landscape, and real process risk.

01

Documents & permissions

DMS, file systems, and technical archives are connected with role context intact.

02

Local retrieval

Parsing, chunking, vector indexing, and re-ranking run in the selected environment.

03

Private model

The LLM receives only approved passages and clearly defined answer rules.

04

Verifiable output

Answer, source location, version, and audit event are returned together.

Potential building blocks

vLLM / Ollamapgvector / QdrantOIDC / Entra IDDMS / ERP connector
The final architecture follows data classification and an infrastructure assessment. Air-gapped deployments are possible, but not a blanket promise.

On-premises or private cloud

Run on your hardware, inside your cloud subscription, or as a clearly bounded hybrid architecture.

Source-grounded generation

Retrieval and answer rules are evaluated with your document set and a defined test set.

Preserve existing roles

DMS, ERP, and identity boundaries remain intact; AI does not create shadow permissions.

Measurable and auditable

Quality metrics and audit logs expose source coverage, performance, and failure modes.

Industrial knowledge work

Built for real processes — not a chat demo.

A pilot begins with one bounded workflow, measurable questions, and documents that already create time cost, friction, or operational risk.

USE CASE / 01

Tenders & RFQs

Structure requirements faster

Extract obligations, exclusion criteria, and open questions from large requests — each linked back to the original passage.

Representative review question

Which mandatory requirements are missing from our proposal?

USE CASE / 02

Technical documentation

Make manual knowledge usable

Help service, engineering, and support find relevant procedures across product variants and document versions.

Representative review question

Which maintenance steps apply to product line X?

USE CASE / 03

Quality & compliance

Find evidence consistently

Search policies, test evidence, and internal standards with permissions, version context, and source references intact.

Representative review question

Which evidence covers this customer requirement?

Engagement model

Prove it first. Then integrate.

Every stage produces a clear decision basis. A pilot is only recommended when the data, use case, and expected value align.

01 / 1–2 weeks

Sovereign AI Assessment

A technical and operational decision brief before any implementation begins.

Included scope

  • Use-case and data classification
  • Infrastructure and integration audit
  • Risk, architecture, and pilot plan
02 / typically 4–6 weeksRecommended start

RAG Pilot / PoC

One bounded workflow with real documents, a test set, and measurable answer quality.

Included scope

  • One prioritized document set
  • Isolated search/answer interface
  • Evaluation, source review, and handover

Custom fixed-price proposal

Get a project estimate
03 / scope-based

Enterprise Implementation

Production integration with your identity, data sources, and operating processes.

Included scope

  • DMS/ERP and identity integration
  • Roles, audit logs, and monitoring
  • Operational handover and optional retainer

Custom project proposal

Get a project estimate

The Assessment starts from €1,500 excl. VAT. For the Pilot and Enterprise Implementation, you receive a transparent estimate of scope, timeline, and investment after an introductory call.

How we work

Small, verifiable steps. No months-long black box.

01

Understand the process

We define users, documents, security boundaries, and one measurable target question.

02

Prove the value

The pilot is evaluated against an agreed test set — including failure cases and source coverage.

03

Integrate securely

Only a robust pilot moves into identity, DMS/ERP, and production operations.

Initial fit check

Let us choose the first process that justifies the effort.

In a short introductory call, we assess whether private RAG fits your use case — and which next step makes technical and commercial sense.

Schedule an introductory call

Helpful for the fit check

  • One concrete document process with a visible time or quality problem
  • Sample documents or a realistic description of the data sources
  • Known security, hosting, or permission requirements

No pitch deck required. One real process example is enough.

FileGPT.dev — Sovereign Document AI for Industry