Video unavailable. Open the source below.
NewJev and Kimi Email Classification
A demonstration classifies a 100-email test set with Jev, then asks a larger model to handle uncertain results.
Added to Jevfast
How it uses Jev
Jev classifies emails and provides confidence. The creator routed predictions below 95% confidence to Kimi K3; 31 of 100 messages were escalated.
App workflow
- Classify email
Jev classifies each email as legitimate or fraudulent.
- Check confidence
Compare the prediction confidence with the creator’s 95% threshold.
- Escalate uncertainty
Send predictions below that threshold to Kimi K3.
- Combine results
The creator reports the combined pipeline’s outcome across the 100-email set.
Why it is interesting
The described cascade reserves the larger model for uncertain cases, combining a fast first pass with a confidence-based fallback.