Open-source demo routes email fraud detection via Jev + Kimi K3: 96/100 accurate for $0.07
nutlope · x · 2026-09-20
Developer Nutlope open-sourced jev-fraud, a two-stage fraud email classifier: Jev classifies 100 emails (50 legit, 50 fraud) in 1.42 seconds; predictions at or above 95% confidence are accepted, uncertain ones are routed to Kimi K3 for independent review. The full pipeline scored 96/100 for roughly $0.07.
- Jev runs up to 20 concurrent calls, with a separate queue for up to 25 Kimi reviews
- The UI lets you run live, see which emails get a second opinion, and track accuracy and inference cost
- Confidence threshold is adjustable; per-email inspection and result export supported
A nice reference for the "cheap high-throughput model first, big model fallback" routing pattern; code includes tests, workflows, and config docs.
More from coding & agent
- mcp-proxmox lets you manage Proxmox VE clusters with natural language — modelcontextprotocol · 2026-09-20
- Tronsave MCP Testnet lets agents buy and sell TRON resources via one interface — modelcontextprotocol · 2026-09-20
- Open questions on JEV: model size, calibration reliability, and a future fine-tuning API — vykthur · 2026-09-20
- One astra general directs 7 jev agents in Warcraft 3: 1,000+ LLM calls for ~20 cents — Vjeux · 2026-09-20
- Composite system1+system2 agent harness runs WC3 micro in RL env, open-sourcing soon — Vjeux · 2026-09-20
- eidoverse-video's threejs/webgpu toolchain is portable to your own agent-built games — repligate · 2026-09-20