Private & Sovereign AI
AI models running on infrastructure we operate in Romania. Your data stays inside.
Open-weight models such as Llama, Mistral and Qwen, served with vLLM on our own GPU servers and network (AS56951) in Cluj-Napoca — prompts and documents are not sent to third-party AI providers and are not used to train anyone else’s models.
PRIVATE AI, ON INFRASTRUCTURE WE RUN
Use AI on sensitive data without handing it to someone else.
Contracts, patient files, source code, financial records — processed by models that run where you can point to them.
Most AI projects start with a public API: you send a prompt and a document, a model somewhere answers. That is fine for marketing copy. It is a different conversation when the document is a client contract, a medical report, unreleased source code or a bank statement. Legal and security teams want to know where the data goes, how long it is kept, who can read the logs and whether it ends up in someone’s training set. Often the honest answer is “we are not sure”, and the project stops there.
We remove that uncertainty by running the models ourselves. WeCode operates its own network (AS56951, peering at InterLAN-IX), its own cloud and GPU servers in Cluj-Napoca, Romania. Our team has been building and running production systems since 2013. We deploy open-weight language and embedding models in an environment dedicated to you, connect them to your data, and give you a clear map of what is stored where. This is the foundation for private knowledge assistants, AI agents that work with internal systems and document processing on files that should never leave your control.
Models on our own GPUs
Llama, Mistral, Qwen and other open-weight models, plus embedding and reranking models, served with vLLM or a similar inference server behind an OpenAI-compatible API, so existing tools can connect without rewrites.
Private RAG and fine-tuning
Your documents are indexed in a vector database inside your environment. Where it helps, we fine-tune a model on your own examples, so it learns your terminology and formats. The resulting weights are yours.
Isolated, controlled environments
A separate environment per client, private connectivity over VPN, role-based access, audit logs and backups. You decide who can query the models, what gets logged and how long it is kept.- Open-weight LLMs deployed on dedicated GPU servers
- OpenAI-compatible API endpoint for your applications
- Private RAG over your documents with a vector database
- Fine-tuning on your data, with the weights delivered to you
- Separate environment for each client, no shared indexes
- VPN or private connectivity to your offices and systems
- Single sign-on, roles and audit logs for every request
- Backups of indexes, configuration and model weights
- Evaluation sets that measure quality on your real tasks
- Hybrid routing: private models for sensitive data, cloud APIs for the rest
HOW WE WORK
From a sensitive use case to a private model in production
We start small, measure on your data, then scale what works.
Map the data and the use case
We look at the task, the documents involved, who uses the answers and what rules apply to the data: personal data, trade secrets, client confidentiality, retention.
Choose and test models
We build a small evaluation set from your real examples and compare several open-weight models, and a commercial API as a reference if you allow it, on quality, speed and hardware needs.
Size the infrastructure
Based on the chosen model, expected users and context length, we propose the GPU configuration, storage and network setup, and an isolated environment for your project.
Build the pipeline
Document ingestion, chunking and embeddings, retrieval, prompts, access control and, where it pays off, fine-tuning. Connected to your systems over VPN or private links.
Pilot with real users
A limited group uses it on daily work. We review answers, logs and feedback, fix weak spots and re-run the evaluation set after every change.
Run and improve
We operate the environment: monitoring, updates, backups, model upgrades when better open models appear, and capacity changes as usage grows.STACK AND TRADE-OFFS
What runs where, and what to expect
A typical setup uses vLLM or a similar inference server for language models, a separate service for embedding and reranking models, PostgreSQL with pgvector or a dedicated vector database, and an orchestration layer for retrieval and agents (for example LangGraph). Everything runs on GPU servers in our cloud, inside a network segment reserved for your project, reachable over VPN or a private link. Indexes, configuration and fine-tuned weights are backed up, and file-level backups can go to WeCodeDrive, also hosted in Romania.
There are trade-offs, and we prefer to name them early. Open-weight models are very capable for retrieval, extraction and drafting, but the largest commercial models can still be stronger on complex reasoning. Dedicated GPUs cost money even when idle, so sizing matters. That is why many clients choose a hybrid setup: sensitive documents and personal data stay with private models, while general tasks with no confidential content can use a cloud API. The routing rules are explicit, logged and agreed with you, not left to chance.
Keeping data in Romania and the EU makes the conversation with your data protection officer simpler, but it is one piece of the picture. If your use case may fall under the EU AI Act, we can look at it together through an AI Act review.
FAQ
Questions we hear about private AI
Are open-weight models good enough compared to the big cloud APIs?
For many business tasks, yes: answering questions over your documents, extracting data, classifying, summarising and drafting. For the hardest reasoning tasks the largest commercial models are often still ahead. We test both on your real data before recommending anything, and a hybrid setup is a valid answer.
Where exactly is my data processed?
On servers in infrastructure we operate in Cluj-Napoca, Romania, on our own network (AS56951). In a fully private setup, prompts, documents, embeddings and logs stay there. If you choose a hybrid setup, we document exactly which requests go to an external provider and which never leave.
Is my data used to train models?
Not for anyone else. Your data is only used to fine-tune or index models for your own environment, and only if you ask for it. Fine-tuned weights and indexes belong to your project and are not shared with other clients.
Do we need to buy our own GPU servers?
No. You can rent dedicated GPU capacity in our cloud and scale it up or down as usage becomes clear. If you prefer hardware you own, we can also help you size it and run it on-premise or in colocation.
Does private AI make us compliant with GDPR and the EU AI Act?
It helps, because you control where data is stored, who can access it and what is logged. It does not replace the legal and organisational work. Compliance depends on the use case, the risk category and your processes, which is why we usually pair this with an AI Act review.
Have data that should stay private? Talk to an engineer.
Tell us what you want AI to do and what data it would touch. We will tell you honestly whether a private model fits, and what it would take to run it.