Zero Data Outside the Organization. How LW Bogdanka Deployed AI in Critical Infrastructure in Two Months
There is a moment in every conversation about AI in a regulated sector when enthusiasm meets the sentence: “we cannot send this data anywhere.” For LW Bogdanka — a listed company operating in critical infrastructure and part of the Enea Group — that sentence was not an obstacle to deployment. It was the specification
!Secure AI deployment in LW Bogdanka critical infrastructure
The Problem: Knowledge That Cannot Be Searched Effectively
Bogdanka, like every industrial organization with a long history, manages extensive collections of technical and operational documentation accumulated over many years — from PDF and DOCX files to scans, which accounted for around 80% of the implementation knowledge base. Finding the right information meant slow, manual archive searches. It consumed specialists' time and, more importantly in an industrial environment, increased the risk of missing critical content.
The standard market answer to this problem is semantic search based on a large language model. But popular tools in this class usually run in the cloud, and each of them would mean transferring strategic data outside the organization's perimeter. For a company listed on the Warsaw Stock Exchange and operating in critical infrastructure, that path is closed by definition.
So the challenge was different than usual: not “how to deploy AI,” but how to give employees modern semantic knowledge search while preserving full data sovereignty.
The Solution: RAG Entirely Inside the Client's Infrastructure
We delivered a system based on TWIN:DESK, PROXY:AI and CDF Platform — a Retrieval-Augmented Generation architecture running entirely on-premise.
The knowledge flow is straightforward: raw documentation goes through OCR and normalization, enters a vector knowledge base with strict data separation between departments, and a local LLM inference server generates answers — each one with mandatory citations to source documents.
Two elements of this architecture are especially important because they determine whether the system is suitable for a higher-risk environment.
- Citation as a requirement, not an option. The employee does not receive an answer “on trust” — they receive a synthesis with direct references to specific pages in source documents. In an environment where incorrect information can have operational consequences, verifiability matters more than eloquence. One hundred percent of system answers include source citations.
- Data separation enforced technically, not procedurally. Query filtering at the vector database level prevents unauthorized access to knowledge across teams, access control is integrated with Active Directory, and every system query is written to audit logs. The boundary is not protected by user discipline — it is protected by architecture.
Methodology: Two Months Because the Process Was Measurable
The whole project — from design, installation and integration with the client's IT infrastructure to training and transfer of ownership rights — was completed in two months using the CDF (Cognitive Deployment Framework) methodology.
CDF systematizes the deployment of cognitive systems in higher-risk environments and produces compliance documentation along the way, including materials relevant to the EU AI Act.
It is worth pausing on what transfer of ownership means in practice: the client did not rent a service whose terms can be changed by another party. It received a system it owns, running in its own server room, that does not send a single query outside the organization.
Results
- Two-month deployment.
- One hundred percent alignment with EU AI Act requirements.
- Zero data sent outside the organization.
- One hundred percent of answers with source citations.
- A drastic reduction in time spent manually searching documentation.
- More accurate operational decisions.
- An architecture ready to expand to further departments without loss of performance.
As Paweł Janowski, underground electrical equipment foreman at LW Bogdanka, put it:
> “When defining the project, we were looking for a partner with the technological capabilities to develop a system consistent with our assumptions. […] Thanks to the professionalism and experience of the allclouds team, we achieved our goals in a much shorter time than originally assumed.”
What This Means for Other Organizations
This project shows something part of the market still does not believe: the capabilities of large language models — semantic search, synthesis, work with knowledge — can be launched safely and efficiently in a fully closed on-premise environment.
The constraints of critical infrastructure do not exclude AI. They require an architecture that treats data sovereignty as the starting point, not as an add-on feature.
If your organization faces the same sentence — “we cannot send this data anywhere” — that is not the end of the AI conversation. It is the proper beginning