OpnViking ApS
Local AI.
Verified.
Your data stays in the building. And you know whether the answers are right.
The problem
Sound
familiar?
Shadow AI
Your people already paste internal data into ChatGPT and similar services. Nobody knows exactly what ends up there.
Blocked project
The AI project is stuck because data protection or customer contracts forbid data leaving the building.
Unknown quality
The pilot runs, but nobody can say whether the answers are right or got worse after the last model update.
All three can be solved without your data leaving the building.
The jargon
Anyone can install it.
Tuning it right is different.
Ollama or vLLM is running within an hour. Then the questions start:
- Which model is the best?
- Q4_K_M or Q8?
- 27B or 122B?
- MoE or dense?
- Temperature 0.2 or 0.7?
- top_p, top_k, min_p?
- 32k or 128k context?
- KV cache full?
- Speculative decoding?
- Why did it suddenly get worse?
You don't have to answer them. I choose the model and settings for your task and your hardware, and measure whether it really gets better.
See the setup reviewFrom 2 December 2027. The original deadline of 2 August 2026 was postponed by the Digital Omnibus.
Packages
Four ways
to get started.
Local AI Check
Your use cases, your data, your hardware. I assess what makes sense locally and what it costs.
Setup review
Ollama, vLLM or a ready-made AI server is already running, but the answers don't convince. I review model choice, quantisation, context, sampling and prompts.
In-house pilot
A model in your infrastructure for one real use case, from a single GPU server to a Kubernetes cluster. With access control, logging and cross-checks.
Quality in operation
Tests built from your real questions. Every model or prompt change is measured before it goes live.
If the check shows that local AI is not worth it for you, I will tell you.
All prices excl. VAT. Hardware is not included.
Configurations
Turnkey,
hardware included.
Servers from Wortmann in Germany. I set them up, check the quality and look after operations.
Compact
One server with one GPU for an internal chat AI and search across your documents.
- Wortmann server with GPU
- Chat interface with company login, reduced to what your users need
- Connection to your documents
- Cross-checked answers
- Setup and training
Company
Several GPUs for more concurrent users, several models and integration with your business systems.
- Everything in Compact
- Several models in parallel
- Permissions per department
- Monthly quality report
Data centre
Several servers on Kubernetes, highly available and fully offline if required.
- Everything in Company
- High availability on Kubernetes
- Offline operation (air gap)
- Documentation for audits
Why there are no hardware prices here: memory and SSDs are scarce right now and prices change week by week. You get a binding quote at today's price.
Industries
Where I
come from.
More than 20 years of infrastructure in these sectors. I know the rules before we talk about AI.
Energy & utilities
Grid and plant data, critical-infrastructure rules, segregated networks.
Public sector
Citizen and personnel data, procurement rules and the wish to be independent of US providers.
Manufacturing
Design data, production know-how and documentation that should not leave the plant.
Logistics
Customer, shipment and contract data across many sites.
Who builds this
No slides.
Running systems.
I am not an AI strategist. I am a platform engineer with more than 20 years of Linux and Kubernetes in energy, logistics, government and manufacturing.
What I recommend runs on my own systems first. When a framework has a memory bug, I find it and report it. When a model starts to drift, I want to see it before anyone else notices.
Powered by Froala Editor
Own systems
In daily use.
Kvasir
My test bench. I use it to test models and settings before they reach you and to measure quality in operation. On request, it can also become a solution tailored to your company.
NOMI
Private AI on my own hardware, with memory and voice. Any behaviour it proposes itself is approved by a human.
SiliconTrailer
My app studio. Native apps for Mac and iOS that keep their data on the device.
Questions
Common
questions.
Which model is the best?
There is no best model. There is the right one for your task and your hardware, and that only shows with your own questions. That is why I test several candidates against a test set before recommending one.
Is the quality as good as ChatGPT?
For most office tasks such as summaries, document search and drafts, current open models are good enough. For very demanding reasoning the big cloud models are still ahead. In the check I test this with your real tasks before you buy hardware.
How do we know the answers are right?
Answers can be cross-checked by a second model, and a fixed test set built from your own questions measures quality after every change. You see the result in the monthly report.
Do we get Kvasir?
Not as an off-the-shelf product. Your staff get a lean chat interface with prepared assistants, without model selection or a jungle of settings. Kvasir is my tool for checking what runs at your site. If you need more, I build a solution tailored to you on that basis.
Why are there no hardware prices here?
Memory and SSDs have been scarce since autumn 2025; some prices have multiplied and they keep changing. A price on this page would soon be wrong. You get a quote at today's price through Wortmann.
Does the server need an internet connection?
No. In operation everything runs inside your network. New models and updates can also be installed offline.
What happens when better models come out?
A new model runs through your test set first. Only if it scores better do I install it, after your approval.
We don't run Kubernetes. Does that matter?
No. For most companies a single server is enough. Kubernetes only pays off with several servers or strict availability requirements.
Contact
Let's talk about
your data.
30 minutes, free, with a live demo. Book directly or through your staffing partner. I reply in English, Danish or German.
Christopher Rau
Founder of OpnViking, Linux and Kubernetes engineer for more than 20 years