Nh1816 Verzekeringen
Full-timeOudkarspel, Netherlands · Apr 2025 – Present
AI Engineer
Nov 2025 – PresentThe first role where I own an AI system end to end rather than one feature of it. The hard part turned out to be everything around the model: keeping the context an agent retrieves accurate, proving output quality with an evaluation set instead of a convincing demo, and designing for the fact that an insurance advisor is not allowed to be wrong. I learned to treat prompts and agent graphs as software — versioned, monitored and tested like the rest of the stack.
- Design and build a multi-platform AI advisory chatbot on Microsoft Foundry, driven by a multi-agent workflow, as part of a Scrum team.
- Set up secure MCP servers and vector stores that let agents query customer and policy data on demand for more accurate, context-aware advice.
- Implemented monitoring of agent interactions, enabling continuous feedback loops and targeted optimisation of system prompts.
- Built an event-driven pipeline that automatically analyses and assesses insurance claims across multiple products, saving the equivalent of 2 FTE.
- Established an evaluation framework that measures the quality of agent output using a golden set and LLM-as-a-judge.
Backend Developer
Apr 2025 – Nov 2025Where I found out what "business-critical" actually means. AFD 2 is a national standard with no room for interpretation, so most of the work was reading specifications precisely, modelling the data exactly as prescribed, and building APIs that external partners could integrate against without needing a support call. It was also my first real exposure to API gateways, security and the human side of onboarding a partner.
- Developed REST APIs for large-scale portfolio transfers compliant with AFD 2, the national Dutch insurance data-exchange standard, within a Scrum team.
- Configured Azure API Management as a secure gateway and handled technical onboarding of external partners, contributing to 15% premium growth.