JEV email classification demo

How one Jev call can answer three different kinds of question about an email at once: whether it is spam (Noul, probability scored against the dataset's labels), which category it belongs to (Choice among marketing, finance, sales, customer support and other) and how urgent it is (Score on a five-level rubric).

npm install then npm run dev, and open http://localhost:3000. Add your TypeSafe key in Settings (the gear icon), which goes only to the app's own server route before being forwarded to api.typesafe.ai; locally you can instead put TYPESAFE_API_KEY in .env.local.

The public Enron-Spam dataset via Hugging Face (SetFit/enron_spam), cited in the README as Metsis, Androutsopoulos and Paliouras, Spam Filtering with Naive Bayes — Which Naive Bayes?, 2006. The Run on dataset view classifies a balanced sample from the test split shipped in the repo.

Yes. The Example tab shows a saved run for visitors without a key, but the README warns that the shipped data/example-run.json is clearly labelled illustrative placeholder data, not Jev output, until someone replaces it with a real recorded run via Save as example.

Category, confidence and probability per option for the Choice question; a fractional score plus probability per level for the urgency rubric; and a 0–1 spam probability for the Noul question. The app also shows the exact request and response trace, question by question.

The repo ships Netlify settings (build npm run build, Node 22, Next.js runtime) and no base directory is needed. The README warns not to set TYPESAFE_API_KEY on a public site unless you also set ALLOW_SERVER_KEY=true on purpose, because every visitor would then be using your key.

Yes — python/ contains a stand-alone script using the same questions as the app. With httpx installed and TYPESAFE_API_KEY exported, python jev_spam_demo.py -n 50 runs 50 spam plus 50 ham and writes metrics, CSV and an HTML report; the --report flag rebuilds the HTML without calling Jev.