On-premise AI server
Hardware selection (GPU, VRAM, RAM, cooling) for the models you need; Proxmox with GPU passthrough, running models (e.g. Qwen, Gemma) in Ollama, llama.cpp or vLLM.
AI in your company – local models and assistants
I work with AI every day: Claude, Codex, Cursor and OpenRouter speed up my code, infrastructure-as-code and reviews. But client contracts, personal data or security logs should not go to an external API. So for sensitive work I use models running locally – and I deploy the same in companies: private AI on your own hardware where data must stay inside, and cloud models where that is safe and cost-effective.
What you get
Private chat for staff
A ChatGPT-style interface (e.g. Open WebUI) on your server, company login (SSO / LDAP), conversation history stays in-house.
AI gateway: local or cloud
One API for all tools (e.g. LiteLLM): sensitive requests go to local models, the rest to Claude / GPT / OpenRouter. Limits, costs and logs in one place.
AI assistants for the IT team
Rolling out Claude Code, Codex or Cursor in a team: rules on what not to send, repository setup, integration with CI/CD and code review. Based on how I use them myself.
AI on company documents
Search and answers based on documentation, procedures or a knowledge base (RAG) – entirely on your infrastructure.
Security and compliance
AI usage policy, tool register, request logs, access control and monitoring (including Wazuh) – for GDPR and NIS2 requirements.
Technologies
- Claude / Claude Code
- OpenAI Codex
- Cursor
- OpenRouter
- Qwen
- Gemma
- Ollama
- llama.cpp
- vLLM
- Open WebUI
- LiteLLM
- Proxmox (GPU passthrough)
- Docker
How we work together
- 01
Use-case conversation
What AI should be used for, what data it will process, how many users. That decides what runs locally and what in the cloud.
- 02
Model and hardware selection
Testing a few models on your examples, a concrete hardware spec with a quote – or a rented GPU server option.
- 03
Rollout
Server, models, chat, API gateway, SSO, monitoring – in stages, with the first users after a few days.
- 04
Maintenance and growth
Model updates (new ones arrive every few months), tuning, new use cases, user support.
Frequently asked questions
Does data really stay inside the company?
With a local model – yes: the model runs on your server with no external API call. The gateway makes sure requests marked as sensitive never reach cloud models.
Is a local model as good as Claude or GPT?
For many tasks – summaries, classification, answers from documents, simple scripts – open models are more than enough. For the hardest coding and reasoning, the best cloud models still lead. That is why I usually propose a hybrid setup.
What hardware do I need?
It depends on model size and the number of users. Smaller models run on a single card with 16–24 GB of memory, larger ones need more VRAM or several cards. After testing on your examples I give a concrete spec – sometimes renting a GPU server turns out better.
Do you use AI in your own work?
Every day. Claude, Codex and Cursor for code, Ansible, Terraform and reviews; OpenRouter for comparing models; local models when I work with sensitive data. So I deploy things proven in practice, not from slide decks.
Have a project or an infrastructure problem?
Write a few sentences about what you want to achieve. I usually reply within 24 hours and tell you straight away whether and how I can help.
Other things I can help with