<feed xmlns="http://www.w3.org/2005/Atom"> <id>https://blog.ruudjuffermans.nl/</id><title>Ruud Juffermans</title><subtitle>Data engineer writing about warehouses, streaming pipelines and AI systems that only claim what they can prove.</subtitle> <updated>2026-11-15T15:46:50+01:00</updated> <author> <name>Ruud Juffermans</name> <uri>https://blog.ruudjuffermans.nl/</uri> </author><link rel="self" type="application/atom+xml" href="https://blog.ruudjuffermans.nl/feed.xml"/><link rel="alternate" type="text/html" hreflang="en" href="https://blog.ruudjuffermans.nl/"/> <generator uri="https://jekyllrb.com/" version="4.4.1">Jekyll</generator> <rights> © 2026 Ruud Juffermans </rights> <icon>/assets/img/favicons/favicon.ico</icon> <logo>/assets/img/favicons/favicon-96x96.png</logo> <entry><title>SCD type 2 in dbt with window functions only</title><link href="https://blog.ruudjuffermans.nl/warehouse-scd2-with-window-functions/" rel="alternate" type="text/html" title="SCD type 2 in dbt with window functions only" /><published>2026-09-13T09:00:00+02:00</published> <updated>2026-09-13T09:00:00+02:00</updated> <id>https://blog.ruudjuffermans.nl/warehouse-scd2-with-window-functions/</id> <content type="text/html" src="https://blog.ruudjuffermans.nl/warehouse-scd2-with-window-functions/" /> <author> <name>Ruud Juffermans</name> </author> <category term="Data Engineering" /> <summary>Dutch municipalities merge almost every year, and a warehouse that only knows the current list quietly rewrites the past. Here is how I versioned the municipality dimension in nl-vehicle-warehouse with lag(), a running sum and a group by — no dbt snapshots, no macros, and a test that guards the one assumption the whole derivation rests on.</summary> </entry> <entry><title>One licence-plate range, three datasets: reproducible sampling for a warehouse</title><link href="https://blog.ruudjuffermans.nl/warehouse-reproducible-licence-plate-sampling/" rel="alternate" type="text/html" title="One licence-plate range, three datasets: reproducible sampling for a warehouse" /><published>2026-09-13T09:00:00+02:00</published> <updated>2026-09-13T09:00:00+02:00</updated> <id>https://blog.ruudjuffermans.nl/warehouse-reproducible-licence-plate-sampling/</id> <content type="text/html" src="https://blog.ruudjuffermans.nl/warehouse-reproducible-licence-plate-sampling/" /> <author> <name>Ruud Juffermans</name> </author> <category term="Data Engineering" /> <summary>The RDW publishes roughly 58 million rows about Dutch vehicles, spread over three datasets that all key on the licence plate. Sample each one on its own and every referential-integrity test in the warehouse fails. This is the sampling rule I settled on instead: one sorted plate range, shared by every dataset, written to a manifest, and reused in CI with one row added on purpose.</summary> </entry> <entry><title>A human-in-the-loop agent with LangGraph interrupt()</title><link href="https://blog.ruudjuffermans.nl/twin-human-in-the-loop-langgraph-interrupt/" rel="alternate" type="text/html" title="A human-in-the-loop agent with LangGraph interrupt()" /><published>2026-09-13T09:00:00+02:00</published> <updated>2026-09-13T09:00:00+02:00</updated> <id>https://blog.ruudjuffermans.nl/twin-human-in-the-loop-langgraph-interrupt/</id> <content type="text/html" src="https://blog.ruudjuffermans.nl/twin-human-in-the-loop-langgraph-interrupt/" /> <author> <name>Ruud Juffermans</name> </author> <category term="Artificial Intelligence" /> <summary>My digital twin has one tool that acts on my behalf: drafting a contact message. It never runs without my decision. This is the mechanism end to end — a graph node that calls interrupt() and checkpoints the thread, an SSE stream that ends with a 'queued for review' notice, and an admin endpoint that resumes the same thread and lets the model finish. With the code on both sides of the pause.</summary> </entry> <entry><title>Cosine similarity is not a relevance score</title><link href="https://blog.ruudjuffermans.nl/rag-cosine-is-not-relevance/" rel="alternate" type="text/html" title="Cosine similarity is not a relevance score" /><published>2026-09-13T09:00:00+02:00</published> <updated>2026-09-13T09:00:00+02:00</updated> <id>https://blog.ruudjuffermans.nl/rag-cosine-is-not-relevance/</id> <content type="text/html" src="https://blog.ruudjuffermans.nl/rag-cosine-is-not-relevance/" /> <author> <name>Ruud Juffermans</name> </author> <category term="Artificial Intelligence" /> <summary>My RAG system over Dutch court rulings refuses to answer when the corpus has no relevant passage. That refusal used to be a cosine threshold, and it did not work: an unrelated question still scored 0.6 against some legal paragraph. A cross-encoder over the top 12 candidates fixed it. Here is why ranking well and being calibrated are two different things.</summary> </entry> <entry><title>No source, no answer: enforcing citations in code, not in the prompt</title><link href="https://blog.ruudjuffermans.nl/rag-citations-enforced-in-code/" rel="alternate" type="text/html" title="No source, no answer: enforcing citations in code, not in the prompt" /><published>2026-09-13T09:00:00+02:00</published> <updated>2026-09-13T09:00:00+02:00</updated> <id>https://blog.ruudjuffermans.nl/rag-citations-enforced-in-code/</id> <content type="text/html" src="https://blog.ruudjuffermans.nl/rag-citations-enforced-in-code/" /> <author> <name>Ruud Juffermans</name> </author> <category term="Artificial Intelligence" /> <summary>My RAG system over Dutch court rulings tells the model to cite every claim. I don't trust it to. After generation every line is parsed and checked against the passages that were actually supplied; what fails is rejected and shown to the user, not hidden. That makes source correctness a design property, and turns the honest metric into something else.</summary> </entry> </feed>
