The Best Triage Builds in the Directory

Triage is the least glamorous thing people build with Jev, the decision model from TypeSafe AI, and the most likely to still be running next year. These are the Jev triage examples we'd point a colleague at: ticket, email, log and document builds, picked because each one teaches something. The pattern itself (closed labels, confidence gates, a human lane) is laid out in the support triage guide, so this page stays a tour.

Numbers are as reported by each builder, with the entry linked. Nothing here is independently benchmarked.

Email builds: the proof of economics

Email is where triage started, and where the receipts are cleanest.

• 500 emails for 3.5 cents. Riley Brown's run, done in seconds, as reported. The number that made half the internet try it on their own inbox.
• Triage across 1,500 emails. Ryan Vogel sorted roughly 1,500 emails by what each one needed. Sorting by required action beats sorting by topic, and this build shows why.
• JevMail. Read-only Gmail triage that stores messages locally and asks for bounded category, urgency and human-sender probabilities. Read-only is the right default for anything touching a real inbox.
• Inbox Zero and Gmail inbox triage show the same idea landing inside existing email tools rather than as a demo.

The deep dive on labels, folders and the cheap-first architecture is the email classification guide.

Builds with an escalation lane

The best triage builds share one habit: they know what to do when the model isn't sure.

• Fraud detection with Jev and Kimi K3. 100 emails in 1.42 seconds, unsure cases sent to a larger model, 96 of 100 correct for about $0.07, as reported. The most-cited cascade in the catalog for good reason.
• Jev Anti-Spam Bot. A Telegram bot that deletes only the spam Jev is sure about. One sentence of design, most of the value.
• firehose-judge. The live Bluesky firehose judged post by post, around $0.00003 a post as reported, with an explicit lane for humans.
• Watching 25 customer groups on WhatsApp. Jev decides whether something needs attention; only then does an LLM write the alert. That's the decide-then-generate split done cleanly.

Tickets, logs and alerts

Same machine, pointed at operations.

• Support answers in a Mac app. Jev reads the question against the built-in manual and picks the article that answers it, or decides none does. The app shows its own docs, so nothing is invented. Reported: 42 of 42 on a held-out set, 0.93 seconds median. Small set, honest framing, great pattern.
• jev-logtriage. Scores log batches for noise, severity and whether an operator should act; code maps answers to suppress, watch, review, notify or page. Nothing is executed on the verdict alone.
• triagedy does security alert triage as a Unix filter, and jevlogs triages OpenTelemetry logs before more expensive analysis.

For the log-grouping trick that turned 22.8 million lines into one question per pattern, see Tocsin.

Documents, code and people

• 900 images in 40 seconds. OCR turns each image into text and Jev picks the category, as reported. The text-first workaround in one build.
• tax-doc-classifier. Classifies PDF pages into IRS form candidates behind a confidence gate, and says outright it is document triage, not tax advice.
• Jev PR Labeler. Labels pull requests, with size measured as scope rather than lines.
• Alan AI application review console. 3,518 internship applications narrowed to 20 in under five minutes, as reported, then handed to humans. Hiring is high-stakes: the model shortlists, people decide, and every verdict should be logged for audit.
• Deal-review router. Maps what comes next for a deal while people make the call.

The thread running through all of them: code owns the actions, Jev owns the label, a human or a bigger model owns the uncertain slice.

Frequently asked questions

What is the best Jev triage example to copy first?

The fraud cascade, because it shows the full shape: fast verdicts, an escalation path for low-confidence cases, and a reported 96 of 100 correct at about seven cents. Pair it with the support triage guide for the design rules.

Can I point these builds at my live inbox?

Several are built to, but start read-only and label-only, as JevMail does. Let any action that can't be undone wait for a human or a second check, never a lone verdict.

Do triage builds need training data?

None of the builds listed here describe a training step; they rely on well-framed questions and confidence thresholds. You still want a few hundred hand-labeled items to check agreement, per the email classification guide.

Is Jev safe for hiring or financial triage?

Only as a shortlisting aid with humans making the decision and verdicts logged. The application console above hands its 20 picks to people, which is the right posture; nothing here is legal advice.