Public Sector
2025
2 Months
by
Benjamin Zamble
The Catalyst
Before ZAGR was founded, this initiative took place during a summer internship as part of a broader project team on the Kvalinn project at Skatteetaten. The objective was to safely process and analyze unstructured public-sector data at scale, requiring the handling of a massive dataset consisting of 218,000+ citizen feedback records.
Technology & Methodology
The team built a multi-stage NLP pipeline to structure the open-text feedback: responses were embedded and clustered using SBERT, PCA, UMAP, and HDBSCAN to surface recurring issue groups within each user task, without manual review or predefined categories. GPT-4o-mini, deployed via Azure OpenAI Service, then generated cluster summaries, titles, and suggested action points - with prompt design and model parameters tuned to reduce hallucination and improve reliability. The result, Kvalinn, gave Skatteetaten's product teams a way to ask not just what users struggled with, but why.
The Questions That Followed
Building Kvalinn meant working directly in the infrastructure most teams only see the output of - structuring prompts, tuning model parameters, and watching a language model's output shape a decision-making tool in real time. It was also a first real exposure to the sandbox-and-agent side of AI development: working inside VS Code, iterating on prompt-driven pipelines, and researching how autonomous AI agents were beginning to be used elsewhere - even though Kvalinn itself was not agentic.
That exposure raised questions that went beyond the scope of the internship itself: how were organisations outside a structure like Skatteetaten's approaching the same problem? Did they have equally strong information security practices? What systems and vendors were they actually using, and who was advising them on the risk? Public-sector AI work tends to sit inside established governance structures almost by default - most organisations don't have that starting point. Following that question is what led directly to founding ZAGR.
Present Expertise
Today, ZAGR applies that same structural thinking - how a system is built, where its outputs come from, and what could go wrong - to AI governance work for organisations navigating the EU AI Act, ISO/IEC 42001, and related compliance frameworks. Current credentials: IAPP AIGP; ISO/IEC 42001 and ISO/IEC 27001 PECB Provisional Auditor.
