I build AI systems for production: assistants that answer with citations, say honestly what they don't know, and only act where a person has approved it — with evaluation guarding every change.
At Politie Noord-Holland I applied AI within criminal law, an environment where privacy and information security are not an afterthought. Most AI projects fail on the data, not the model; so I build the foundation first and the assistant second.
What a team can expect from me
01
Sources behind every claim
An answer without a retrieved passage is no answer. Every claim points to the document or ruling it came from.
02
Guardrails before the model
Prompt injection, PII and off-topic questions are caught in plain code before a single token is spent. Fail closed, never crash.
03
Human in the loop
Actions with consequences — sending a message, changing something — wait for approval. A hard daily cost limit keeps the bill predictable.
04
Evaluation on every change
Labelled datasets and thresholds in CI: retrieval recall, correct tool selection, task completion. A regression blocks the merge.
Where I did this
2024
Data Analyst
Politie Noord-Holland
Applied AI and automation within criminal law — where confidential and privacy-sensitive data is the norm and every outcome has to be traceable.
2023
Data Engineer
Netherlands Aerospace Centre (NLR)
Built the data foundation AI applications have to stand on: clean, modelled and tested data in a research environment.
Since 2020
Instructor
Udemy
Explaining complex technology to people without a technical background, 241 students with a 4.2 rating — including what AI can and cannot do.
No source, no answer.
Guards in code, not in the prompt.
Risky actions only with human approval.
Projects that show it
Two own AI systems running in production — with guardrails, evaluation and the code in the open.
Digital twin: an AI agent that answers for me — with sources
A LangGraph agent over my CV, projects and writing that answers visitors' questions in a streaming chat: every claim cited, refusal when there is no source, a hard daily cost cap, and human approval before it acts on my behalf.
Strafrecht-RAG: no source, no answer
RAG over 1,906 Dutch criminal-law rulings with one rule: no source, no answer. Every claim ends in an ECLI, section and paragraph — and that reference is checked in code, not requested from the model.
AI you'd dare to show a client?
I'm looking for a team where I can do this work properly. A conversation is easily arranged.