For the Ministerie van KGG we built a sovereign AI assistant on the basis of our platform soev.ai. The organisation works daily with extensive, fragmented and changing laws and regulations, technical requirements and standards. Unlocking that knowledge quickly and reliably saves a great deal of time. The assistant supports staff in finding relevant regulations and requirements, and in translating them into concrete steps in the work process.
We developed the solution together with two Dutch partners: Intermax and UbiOps. The full solution, both the application and the AI layer, runs in a separate environment within the Dutch private cloud of Intermax. UbiOps handles the model orchestration and the GPU optimisation. The language models used (gpt-oss-120B and mistral-large) are open-weights models with an Apache 2.0 licence that run entirely on our own infrastructure. That way the organisation stays independent of foreign cloud providers and keeps a grip on data and model.
The assistant is built on reliability and transparency. A multi-agent flow matches the different types of search question, and every answer points to the sources it is based on. Visible reasoning steps, citations and knowledge filters make an answer verifiable before a member of staff acts on it. People decide, the system substantiates.
We tested the solution both quantitatively and qualitatively. With the RAGAS framework, an LLM-as-a-judge approach, we measured a context recall of 0.89 and a faithfulness of 0.96. That means the retriever fetches the large majority of the relevant information and that the answers are based almost exclusively on the sources provided, with very little invented information. In user experiments, participants rated transparency at an average of 4.2 out of 5, and the degree to which the assistant made up information at 2.2 out of 5 (lower is better).
Those same experiments produced an estimated time saving of about an hour per research question, against a current effort of around two hours. Participants scored the time saving at an average of 4.4 out of 5. A financial analysis showed that the net value scales above all with wider deployment across the organisation, with the cost per user falling as more teams use the assistant.
Together with the organisation’s CISO we laid a foundation for security and compliance: a BIO and a DPIA quick scan, automated security scans (internet.nl, Google Lighthouse and the OWASP Top 10 via Semgrep and dependency checks) and an architecture description following the C4 model with an accompanying risk analysis. The main next steps are GraphRAG to follow references between documents and answer broad questions more completely, further data isolation so we can maintain the application without access to the data, and migration to the organisation’s SSO.
