Build log. I published my first n8n node.

1pm, the idea: put Jev, TypeSafe's decision model, inside n8n.

Jev does not write text. It reads a field and answers your question with a probability for every option.

First thing, check whether it already existed. It did: six public packages. I read the code of all six and noted what none of them did. I found a gap. Nobody measures the confidence threshold. They all ask you to pick a number between 0 and 1 and hope.

That gap became the third operation in my node, calibration. You give it labeled examples, it returns accuracy and coverage at every threshold and recommends one. In my test, 0.5 let one wrong answer through and 0.7 let none.

Before writing the node, I ran a batch of tests on the model. Three things I learned:

• the same call repeated returns different probabilities, and no decision above 0.95 changed between runs • 24 items in one request give the same answers as 1 item per request, at half the cost • 28k tokens of unrelated text in the state changed nothing. 2k tokens of text that looked like one of the options turned two answers wrong

23 unit tests, one live run against the API, three workflows executed inside n8n. By night my first contribution to the n8n community was live.

Over the next few days I will show it applied: lead triage, extraction checks, content filtering.

Video: https://www.youtube.com/watch?v=0hxoySIlnyw&t=2s Package: https://www.npmjs.com/package/n8n-nodes-jev-gate

#n8n #automation #ai #jev