Play the video in the post belowPost on X
- Views
- 59,776
- Likes
- 870
- Reposts
- 52
Read from X on . Counts change daily.
Hassan tested a two-stage pipeline for email fraud detection: Jev classifies every email first, and only the uncertain ones get a second look from a larger language model, Kimi K3. He fed it 100 test emails, a mix of 50 legitimate and 50 fraudulent, and Jev classified all of them in 1.42 seconds.
The part he calls underrated is that Jev returns a confidence score alongside each classification, so his code could route anything under 95% confidence to Kimi K3 for a second opinion. 31 of the 100 emails fell below that threshold. After Kimi K3 reviewed those, the combined pipeline reached 96 correct out of 100, in 16 seconds total, for about $0.07: $0.068 from Kimi K3 on Together AI and $0.003 from Jev. He calls it an interesting pattern: use a fast specialized model like Jev for the narrow task, then route the uncertain cases to a larger LLM, which he thinks suits fraud checks or anything real-time.
