Gijs
Jansen

I took four dbt tests written as English sentences off my laptop and onto Databricks. The schema.yml didn't change. Everything underneath it did, and the 50,000-review production run failed its own gate by one point of recall, which turned out to be the most interesting part.

Most LLM features in data tools generate text you then have to parse. I built five demos around a model that returns typed judgments instead (probabilities, choices, scores), so English becomes a column you can filter, test and sort.

not_null and accepted_values will happily pass a customer called Mickey Mouse, a return coded late whose comment says the box arrived crushed, and a ticket with someone's IBAN in it. So I wrote four dbt tests as English sentences and scored them against an answer key.

A pipeline can run successfully and still be wrong. The two failures that matter most are silent by nature — the data stopped arriving, or it quietly changed shape — and neither shows up in a green DAG. Here's what to measure, and how lineage turns an alarm into an assignment.

A stream of events isn't sessions until you decide where one visit ends and the next begins. You can do that in three window functions — no self-joins, no Python — and the whole thing hinges on the same frame-clause detail that trips people up on running totals.

The schema glue is gone, but most teams still hand-write the other half of every API pipeline — pagination loops, auth headers, and "what changed since last time." dlt's REST API source turns all of it into configuration. Here's the declarative shape, end to end.