Two-Stage Fraud Detection Funnel (Jev + Kimi K3)

Overview & result

Hassan’s demo first classifies 100 emails with Jev, then sends uncertain cases to Kimi K3. The linked post reports a 1.42-second Jev pass, 96 of 100 final answers correct and about $0.07 in total cost. The video shows a live run.

From the original post

Jev + Kimi K3 for fraud detection!

TLDR: Jev classified 100 emails in 1.42 seconds, then I routed the uncertain cases to Kimi K3. The full pipeline got 96/100 correct for only ~$0.07.

Video is not sped up, check out the live run!

Here was my process:

I gave Jev 100 emails to https://t.co/A9pPb6dFMy

How Jev fits in the loop

1. Classify the email batch with Jev.
2. Route uncertain cases to Kimi K3.
3. Compare the final results with the 100-email test set described by the author.

How to reproduce

1. Watch the linked live run and read the author’s process in the original thread.
2. Re-run on an independently labeled email set before applying any threshold in production.

Why this build matters

It illustrates a two-stage decision flow that reserves a slower model for ambiguous inputs.

Reported performance

Cost: $0.07 / 100 emails

Latency: 1.42s / 100 items (14ms / item)

Limitations

• The linked post reports a single 100-email demonstration, not a production fraud benchmark.

Patterns