Jev demos — datagobes

Five small, self-contained demos of Jev inside tools data engineers already use: 01 Cringe-o-Meter, 02 semsql, 03 VISUALIZE, 04 dbt semantic tests, and 05 the same dbt tests on Databricks. Each demo lives in its own folder with its own README, dependencies and instructions, and the MIT license covers every one.

It brings semantic SQL to DuckDB: plain-English judgments like jev_noul(body, 'asks for a refund') > 0.8 become typed columns you can filter, sort and aggregate. The judgment is the column, so text classification becomes ordinary SQL.

Yes. Most demos run without an API key in a clearly labelled SIMULATED mode. For live results, put TYPESAFE_API_KEY=... in the demo folder's .env. The README stresses that live versus simulated is always labelled.

dbt tests written as English sentences catch rows that pass every structural test but are still wrong or leak PII. The demo scores them against a hidden answer key and a regex baseline, on both a local setup and Databricks.

Jev runs in a Unity Catalog Python function that reads the key from a UC secret; judgments are cached in Delta, and a 50,000-review production run is scored against planted star flips and a blind audit.

A SQL query ends in VISUALIZE '<intent>'; code proposes only valid charts, Jev scores which one answers the intent, and code builds the chart or dashboard. Chart-building stays deterministic — Jev only ranks the valid candidates.

They are built by DataGobes and released under the MIT license, which covers every demo in the repository. A CHANGELOG.md tracks what changed and when, and the README notes that numbers come from logged runs.