The Weirdest Jev Builds (Complimentary)

"Weird" is a compliment here. When a new primitive shows up, the strange experiments are how everyone finds out what it's for, and Jev, TypeSafe AI's decision model, attracted a lot of them. This is a tour of the weird AI projects in the directory's tools and apps shelf, the unusual Jev builds that made us laugh first and take notes second.

The tidy, category-by-category version is the use-cases tour. This one is sorted by eyebrow height. Numbers, where they exist, are as reported by the builders.

Weird AI projects that are secretly serious

The best weird builds are only weird for about a week, until everyone copies them.

• pg-jev and JevQL. Semantic questions inside Postgres: a WHERE clause that asks a model about each row. pg-jev is an extension; JevQL does it in vanilla Postgres with no extension. Sounds like a party trick until you've written your fifth ILIKE for "angry customers." Why it works, and where it shouldn't be used, is the natural-language database queries piece.
• fast-jev-compaction. A Claude Code plugin that scores tool-call pairs for deletion or truncation, so a coding agent's context gets pruned by judgment instead of by age. Explained properly in what fast-jev-compaction is.
• invalidate. Every fact an agent remembers gets a lease, and Jev decides whether new evidence ends it. Memory that goes stale on purpose. We have met humans who need this.
• Port Cleanup. A macOS utility that tells you which listening ports are safe to stop. Tiny, specific, instantly useful.

Weird on purpose

No business case was harmed in the making of these.

• Jevinci. Paintings made from Jev's pixel probabilities. Nobody asked; we're glad someone did it.
• A city of 79 Jevs. An architect Jev places houses, and each resident asks the model what to do next: sleep, work, eat, wander.
• Music intelligence audit. 3,274 songs across 114 genres scored for "Soul vs Subversion," for about $0.0001 in total, as reported. The price of settling a music argument is now effectively zero.
• 1980s expert system on Lisp. An old idea re-run with a decision model in the loop. Retro computing with a point.
• Real-life magic 8 ball. Exactly what it says. A decision model was always going to end up here.

Tools that break their own category

• Jev Calc. Kitze's calculator notebook where you type "2150 euro to usd" and it works out the number.
• CLI did-you-mean wrapper. Mistype "git record" and it suggests "git commit" by meaning, not typo distance.
• jevcumber. Cucumber tests run from the .feature file alone. No step definitions.
• A syntax highlighter on Jev. Code tokenizes, Jev names the language, colours every word and says which of nine lint rules fire. The rules live in code; Jev only answers.
• Typed-to-UI textbox. A text box that turns into whatever UI you type.

None of these generate anything. Every one is a pick from a closed set, dressed up well. That's the whole trick, and why it isn't generation is worth knowing before you try your own.

Ambient weirdness

The strangest category: software that watches and mostly decides to do nothing.

• An always-on assistant with no wake word. Jev listens constantly and uses the probabilities to tell a command from ordinary talk.
• Real-time Clippy. Only wakes up when it thinks you're struggling. Clippy, redeemed.
• UXRay. A heads-up display that asks whether what's on screen is a dark pattern.
• Jev-controlled traffic light. Runs a simulated light on near-real-time traffic feeds for under a cent total, as reported. The builder's highlight was watching it decide to do nothing, which is the most relatable AI behavior in the catalog.

The common thread is restraint. A cheap, fast judge means you can afford to ask "should I act?" constantly, and the right answer is usually no.

Frequently asked questions

What is the weirdest Jev build?

Our vote goes to Jevinci, paintings made from pixel probabilities, with the do-nothing traffic light a close second. Both are in the directory's tools and apps shelf, reachable via search.

Are any of these weird builds actually useful?

Several: semantic Postgres queries, context compaction for coding agents and expiring agent memory are all patterns you could put in production. The use-cases tour shows where they fit.

Can Jev really run inside a database query?

Builders report doing it through pg-jev and JevQL, which ask Jev a question per row. The trade-offs on cost and repeatability are covered in the database queries piece.

How do I get my weird build listed?

Submit it with a link to the source and any numbers you measured. Weird is welcome; receipts are required.