The challenge
An enterprise IT distributor needed to validate practical generative AI use cases that run entirely within its own on-premise hardware environment, without cloud-based LLM services, to satisfy data-sovereignty, security and compliance requirements.
At a glance
ModelsDeepSeek, LLaMA, Mistral (local)
ArchitectureRAG, local vector store, text-to-SQL
DeploymentClient's own on-premise hardware
What we did
SWICON implemented a local RAG architecture with a vector store and a text-to-SQL / text-to-report pipeline on the client's hardware. Contract documents and industrial machine logbooks were ingested, chunked and embedded; structured sample databases were connected for report generation. Every use case allows switching between three locally hosted models (DeepSeek, LLaMA, Mistral).
The result
Four use cases are running live on-premise: natural-language search and Q&A over the contract corpus, natural-language search over equipment and maintenance logbooks, and text-to-report on two structured SQL databases. Delivery took around four weeks.
