What it is
A quiet first reader for a store inbox.
What it does: Each email goes to Jev once with eight questions answered in parallel: topic, how soon a reply is needed, how the sender feels, and five risk checks (health or safety, chargeback or legal, public complaint, refund request, sales pitch). Rules in code then pick a lane. Risk flags trigger early, a pitch must be clear before it is archived, and anything the model is unsure about goes to a person.
Why it belongs in the showcase: A worked example of asymmetric thresholds, with the misses and near-misses written up next to the results.
Caveats: Test inbox of 20 emails; n8n flow not yet run live.
Evidence
Public MIT README; author-recorded run on 2026-10-06 with `jev-1.13.0`, raw output in `results/raw.json`: 19 of 20 test emails in the right lane, 0 urgent or risky emails sent to automation, median 338 ms per email, about 1,000 input tokens per email, about 2.5 US cents for 600 emails a month; safety rules run first and unsure answers go to a person; thresholds were set before the run; a shadow-mode n8n workflow is included.
Caveats
- The 20 emails are a test inbox written for the experiment, not a live store
- The n8n workflow has not yet been run inside a live n8n instance
- On a real inbox, customer emails go to TypeSafe; check the data terms and the store's privacy policy first
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.