TypeSafe Jev — gtaras7

A folder of experiments. The README describes experiments in building software around a decision model rather than around a chat prompt, and says each project is self-contained with its own README, its own tests, and its own measured results. Only one is documented at the root: cv-screen, a local screening workbench for a folder of CVs. No install step or published package is described here.

The repo positions it as a self-contained experiment rather than a hiring system, and cv-screen itself runs only a small fixed set of typed questions per candidate with plain readable arithmetic on top. What it does claim is that the scoring layer is separable from the model calls. Any accuracy numbers would live in the cv-screen README, which the root README does not restate.

That separation is the point of the design. Because the typed Jev answers are kept apart from the arithmetic, editing a weight, a cap, or a question re-scores every stored candidate in about 20 ms and costs nothing, and never re-calls Jev for candidates already judged. Structuring software that way around a decision model is what the repository says it is testing.

Partly, and the repo is specific about where. Each project carries its own tests and measured results, and the cv-screen README documents what it tests about the model, how to run it, the results it produced, and the two bugs its own test data caught. The root README gives none of those numbers, so they are not available without opening the subproject.

MIT, with a LICENSE file at the root. cv-screen is built on TypeSafe's Jev model, so a TypeSafe API key is required to judge candidates at all. There is no published package to install, so you clone the repository and follow the project README.

The CV contents of every candidate go to Jev for judgment, which is why the entry describes it as a screening workbench rather than a hiring system. Nothing in the repository runs a local model or offers an offline path; the API key is what buys the judgments, and the arithmetic on top runs locally.

Either. Point it at a single CV or at a folder of hundreds, and each candidate comes back with one verdict from the fixed set of typed questions, presented as a sortable shortlist. The role being hired for is a policy you edit in the app rather than a config file, which is what makes the re-score cheap.