Sift: AI Email Triage with Jev

Search for intent, not just a word

An email can ask for money back without saying “refund.” It can also mention a refund without requesting one. Sift explores that gap: describe the messages you mean, and Jev evaluates whether each message matches the intent.

Nader Dabit’s demonstration uses 500 synthetic emails. Labels can be combined to narrow the inbox; a separate triage pass evaluates messages along multiple dimensions, with application code turning those judgments into a priority queue.

What to watch in the demo

• Start with one intent, such as finding a request for a refund.
• Inspect the actual messages behind the matches, especially ambiguous wording.
• Compare the keyword baseline and model judgments instead of assuming either is always correct.

A useful next experiment

Try a small, labeled evaluation set for your own inbox workflow before connecting real mail. Include negation, quoted history and borderline requests. Review false positives as carefully as missed messages: both affect which customers get attention.

Know what you are seeing

This is an author-provided project animation, not a connected Gmail account on JevFlow. The original repository documents live and replay modes. JevFlow neither reads your inbox nor sends replies; Try leads to the early-access form preview.