Local AI integration: your own language models with full data sovereignty

Not if you don’t want it to. We rely on local AI: the language model runs on a server in your office, in your server room or, if you prefer, in a Swiss data centre. Inputs, documents and results are never sent to a third-party AI provider. We only integrate cloud models if you explicitly ask for it – and we disclose beforehand which data goes where.

For most business tasks, yes. Current open models answer questions, summarise documents, write texts and use tools at a level that is entirely sufficient for day-to-day work in SMEs. Fine-tuned for your tasks, they can even outperform a general-purpose cloud model in those areas.

Fine-tuning means training an existing open language model further with your own examples – for instance on your technical terms, text templates or typical workflows. At the start, a good base model with access to your knowledge (RAG) is often enough. Fine-tuning pays off when the model needs to master a particular style, format or technical language reliably. We assess this together with you using real examples.

That depends on the model size and the number of simultaneous users. For many SMEs, a single server with a powerful graphics card is enough. In the initial consultation you receive a concrete cost estimate for hardware, set-up and operation. Unlike cloud services, there are no per-request costs.

We do – if you wish. We take care of monitoring, updates, backups and switching to newer models. Instead of an anonymous hotline, you have a dedicated contact person in the Lucerne region who knows your system.

RAG stands for Retrieval Augmented Generation. Put simply, the language model first searches your own documents for relevant passages and then answers on that basis. This gives you answers drawn from your knowledge rather than the model’s general knowledge – with sources cited. With us, this search also runs entirely on your server.

A chatbot answers. An agent acts: it can plan several steps, call systems, check intermediate results and see a task through to completion. Wherever things become critical, we build in an approval step so that a person confirms before anything is triggered.

Yes, if there is a concrete use case. It makes sense wherever text is regularly read, sorted, summarised or transferred today – and especially where confidential data is involved. Without such a use case, we advise against it – AI for its own sake only costs money.

Open language models such as Qwen, Mistral or Llama, run with vLLM, llama.cpp or Ollama on NVIDIA GPU servers. For fine-tuning we use LoRA techniques, and vector databases for searching your own knowledge. Integration with your software is done with Python, LangChain, n8n and our own MCP servers.