Developer tests Jev on 300 emails: clear mail sorts, borderline misfires

The developer built this tool after seeing a video comparing Claude Sonnet, GPT-Luna, Gemini-Flash, and Jev for email classification. Motivated by wanting to separate newsletters and payment records from mail that actually needs attention, they used IMAP to read iCloud mail read-only and sent each message's sender, subject, and first 3000 characters of body to Jev.

The implementation had two versions: classify.py for judgment only, and sort_applemail.py which added actual folder moves. New mailboxes were created for credit card notices, bank notices, postal notices, advertising, and two "other" categories. The tool opened a browser page where the user could pick a period (7 days, 14 days, or 2 months), start classification, then execute the sorting.

Notably, the developer concluded that little was enabled specifically by Jev, and that simply registering folders by email address, or running a local LLM for privacy, might work just as well. The borderline categories — especially "needs reply" — showed the most room for improvement, suggesting that for now, the value of AI-based email sorting hinges on whether those ambiguous rules can be defined clearly enough.