Jev + Kimi K3 cascade classifies 100 fraud emails at 96% accuracy for $0.07
nutlope · x · 2026-09-17
Developer nutlope shares a cascade inference pipeline for email fraud detection: the fast specialized model Jev classified 100 emails (50 legit, 50 fraudulent) in 1.42 seconds, exposing confidence scores. The 31 predictions below 95% confidence were routed to Kimi K3 as a fallback.
The combined pipeline hit 96/100 accuracy in 16 seconds for about $0.07 total — $0.068 from Kimi K3 on Together and just $0.003 from Jev. He argues this pattern — fast narrow model first, big LLM only for uncertain cases — could be a game changer for fraud detection and other realtime use cases.
More from coding & agent
- Alignment drift study: one reward hack raises GPT-5.5's re-hack rate from 10% to 64% — maksym_andr · 2026-09-18
- Typesafe goes from two-year stealth cold start to 140k waitlist signups in under 36 hours — damianplayer · 2026-09-18
- New blog finds data issues in a frontier benchmark, ports Agents' Last Exam CLI subset to Verifiers v1 — dejavucoder · 2026-09-18
- After pushing 2B tokens through DeepSeek V4.1 Flash, dev cuts AI bill from $300 to $30/month — gaganghotra_ · 2026-09-18
- Prompting lesson from long-running agents: telling it 'you may skip' makes it skip — BraceSproul · 2026-09-18
- Don't fine-tune: an MCP over Claude handles your 700-episode podcast archive — StewartalsopIII · 2026-09-18