Jev: AI decisions in your workflows

The Jev nodes add a different kind of AI to your workflows than text-generating models: a decision model. You give it a text and some questions, and it returns structured answers with probabilities (yes/no, one option from a list, a level on a scale). It invents nothing outside what you let it choose, and the workflow keeps control of the flow.

The problem

Many workflow decisions cannot be written as a rule:

• Is the customer’s message a complaint or an enquiry? Which team owns it?
• Does this lead fit your ideal customer? Is it urgent?
• Does the form carry spam, personal data or an attempt to manipulate your assistant?

Until now there were two poor options: keyword rules that fail on any new phrasing, or a generative model — slow, costly and unpredictable — that can also drift from the flow.

The solution

Node	Purpose	Outputs
Jev: Ask	Several questions in one call	1
Jev: Yes/No condition	An If with judgement	Yes · No · Uncertain
Jev: Classify	Pick a category	One per category + Uncertain
Jev: Score	Simple or weighted composite score	Low · Medium · High
Jev: Verify rules	Moderation and validation	Pass · Flagged
Jev: Locate in image	Locate something in an image (experimental)	1

• Fast: most queries take about 100 ms and questions are evaluated in parallel, so asking more costs little latency.
• Cheap: 0.042 $ per million input tokens; output is free. A thousand classifications of a short ticket cost about 2 cents.
• With probabilities: every answer carries its confidence. What is certain proceeds alone; what is doubtful goes down a human review path.
• Your key: you use your own TypeSafe account (typesafe credential); Floogos does not bill that usage.

How to design well

1. Deterministic things in code: dates, amounts and identifiers are checked with regular nodes. Jev only judges.
2. Atomic questions: one property per question; many independent questions in one call.
3. State as an object with named fields and only the relevant context.
4. Contrastive categories: say what each covers and what it does not.
5. Route by confidence: high, act; medium, ask for confirmation; low, a person. Calibrate thresholds with your own cases.

Before you start

• Language: Jev performs best in English; Spanish works with lower accuracy. Test with your data.
• Privacy: the evaluated text is sent to TypeSafe (a third party). Do not send what you cannot share.
• Model version: jev-latest moves to new models with releases; pin a version in production.

Use cases

See the Use cases section in the menu: ticket triage, lead qualification and form moderation.