Six Jev classifiers route 1.2M decisions a day
The author combines six Jev classifiers into one router for reranking, tool-call gating, citation checks, model routing, injection detection and page classification, claiming it is 444x cheaper than the frontier model.
I BUILT JEV. SIX AGENTS, ONE REFLEX. IT ANSWERED 1.2 MILLION QUESTIONS TODAY WITHOUT DECODING A SINGLE TOKEN. 444× CHEAPER THAN THE FRONTIER MODEL IT REPLACED. Classifiers don't generate text. They return a choice, a score, or a probability. Jev is six of them bolted together under one router, deciding things in 70 milliseconds instead of five seconds. Here's who's on the roster: - RANKO reranks search results, scored 21,480 queries today, top-1 relevance up 11% - WARDEN guards every tool call before it fires, held 1 destructive command at p=0.98, "rm -rf ./prod" never ran - CITA checks every citation against the quoted claim, caught 847 fabrications, 99.1% recall on the eval set - ROUTE picks the right model per request, routed 38,820 calls, 71% stayed on the small model - AEGIS screens inbound text for prompt injection, flagged a page that read "ignore previous instructions" at p=0.99 in 66ms - LABEL classifies pages into guide/reference/tutorial/changelog, hit 44,870 docs today without one hallucinated label Peak moment was LABEL running 777 judgments in a single batch, 0.7 seconds, zero tokens decoded. Would have cost $1,840 on the frontier API. Cost $0.0025 here. Daily savings ticker opened at $1,180. Closed at $12,480. The thing people miss is that most "LLM calls" in production aren't generation. They're decisions. Does this match? Is this safe? Which bucket? Who do I send this to? None of those need a 500-token completion. They need one typed answer with a confidence number. I didn't call GPT once today. I didn't burn a frontier credit. Six classifiers did, and every one of them returned under 80 milliseconds.