AI where data cannot leave the organisation: open models in our GPU cloud in Lithuania (EU), Claude through Amazon Bedrock in an EU region, or a model on your own servers. No training on your data, access and logs in your hands, GDPR, NIS2 and DORA compliance.
Built for ISO 27001NIS2 DORA
A clinic has patient records, a law firm has case files, a leasing company has applications with financial statements, a manufacturer has drawings and pricing. AI would help each of them as much as anyone else, but the data cannot go to a public service: customer contracts forbid it, industry regulation forbids it, or common sense does. The usual result is that employees use ChatGPT from personal accounts anyway, only nobody sees it and nobody controls it.
Private AI solves this dilemma with architecture, not promises. We run the model where your data is allowed to be: open models (Llama, Mistral, Qwen, Gemma) in our own GPU cloud in a data centre in Lithuania (EU), Claude through Amazon Bedrock in an EU region in Frankfurt, inside your AWS account, or, when full isolation is required, on your own servers. In no case is your data used to train models, access is governed by the identities in your Microsoft 365 tenant, and every request and agent action is written to a log you can show an auditor.
We choose based on where your data is allowed to be and what quality the task needs. Often two options are combined.
Llama, Mistral, Qwen or Gemma run on NVIDIA GPU servers in our data centre in Lithuania (EU), dedicated to your environment. Data stays in the EU, and the price is fixed per GPU, not per request. Suited to document-based assistants, classification, summaries and document recognition.
When you need the strongest document analysis and agent quality: Claude models run in Amazon Bedrock in the Frankfurt region, inside your AWS account, with no training on your data, with your keys and CloudTrail logs. Data stays in the EU, and you pay by usage.
A fully isolated environment with no internet: a GPU server on your premises, with the model and the vector database inside. It is rarely needed, mostly for defence, critical infrastructure or contractually restricted data, but when it is needed there is no other way. We design, deploy and manage it remotely through your VPN.
The same solutions as on public platforms, except that the model and the data stay under your control.
RAG search across contracts, procedures, case files or records, with a link to the source and respecting your existing access rights.
Invoices, applications, KYC documents and statements are broken down into fields by a private model, without being sent to foreign OCR services.
Multi-step tasks in your systems with least-privilege access, approval steps and a complete action log.
Email, complaints, incidents and feedback are sorted and routed by your rules, and sensitive data does not leave.
Summaries of records, case files, reports and meetings, and drafts of replies and contracts based on your templates.
A hybrid: a private model removes personal data, and only the anonymised text goes to Claude or ChatGPT.
A programming assistant that works with your codebase, for when code cannot leave the company.
A vector database and semantic search across millions of documents, connected to SharePoint, your ERP or a DMS.
A short comparison, which is where we start the conversation.
| ChatGPT / Copilot (public services) | Claude through Amazon Bedrock EU | Open models in the MDP CLOUD GPU cloud | |
|---|---|---|---|
Where the model runs | on the vendor's platform (OpenAI, Microsoft) | in your AWS account, Frankfurt region | in our data centre in Lithuania (EU), in your dedicated environment |
Data used for training | no (in business plans and the API) | no | no |
Data location | EU storage depending on the plan | EU (Germany) | Lithuania |
Model quality | highest | highest | sufficient for most business tasks |
Cost | per user or usage-based | by usage | fixed per GPU, independent of the number of requests |
Full isolation (no internet) | |||
Your keys and logs | partly (administration logs) | ||
Suitable for regulated sectors | limited, depending on the data category |
From a data assessment to a managed environment in 4 to 8 weeks.
Together we list what data will go into the AI, which categories it falls into under GDPR and your industry rules, and who may have access to it. This tells us whether Bedrock in an EU region is enough, whether a model in Lithuania (EU) is needed, or whether full isolation is required.
We select the model by task and language (we test several for your language), design the vector database, the connections to SharePoint, your ERP or a DMS, and the access model based on the identities in your Microsoft 365 tenant.
We build a dedicated GPU environment, enable logging and monitoring, load the documents and test with your staff on historical cases until answer quality and source accuracy meet the agreed threshold.
We manage it like the rest of your infrastructure: we update models and components, monitor performance and cost, take backups and review logs. Once the first solution works, we add others in the same environment.
If you have your own team and want to run models yourselves, we rent NVIDIA GPU servers in a data centre in Lithuania (EU), with management and backups.
For general creative tasks the large commercial models are still ahead, so for those cases we offer Claude through Bedrock in an EU region. For typical business work (search across your own documents with a source, classification, summaries, document recognition) the quality of open models is sufficient, and we verify that with tests on your documents before you sign.
The latest open models understand and write well in most European languages, but they differ from one another, so we test the language with your own texts and choose the best fit. For document recognition we use OCR that handles accented characters.
The cost of a private model is fixed per GPU environment and does not depend on the number of requests, so with heavy use it is often cheaper than a licence for every user. Bedrock is billed by usage. After the first conversation we provide a calculation based on your expected load.
Yes, and this is the main reason to choose private AI: data stays in the EU, is not used for training, access is restricted, actions are logged, and the infrastructure is managed in line with ISO 27001 requirements. We also help prepare the documentation for an audit: a data flow description, a risk assessment and a data processing agreement.
Yes, and we recommend it: one process, one set of documents, one group of users. The first solution usually takes 4 to 8 weeks, and then others are added in the same environment without new infrastructure.
Repetitive tasks that a person does today can run by themselves, more accurately and without waiting.
In a professional services firm, time is the product, and a large share of it goes on Outlook enquiries, proposals, updating HubSpot or Pipedrive…
Bookings arrive through Messenger, reviews land on Google Business Profile and Wolt, sales live in Lightspeed, and supplier invoices reach SAP…
A supplier sends a thousand products with no descriptions for PrestaShop or WooCommerce, customers ask about stock in Lightspeed and parcels with…
This service is delivered by specialists with valid international certifications. Their expertise is confirmed by independent organisations, not by us.
AI systems are deployed with protection against data leakage and model abuse.
GPU infrastructure and AI workloads are operated following NVIDIA practices.
The AI security programme and its controls are managed following the ISACA methodology.
We reply within 1 business day
