Gevva0

Project intro

Gevva0 is a cascading decision engine for local AI: a four-tier pipeline that answers a question at the cheapest tier that can handle it, instead of calling a frontier model for everything. Tier -1 is static routing, Tier 0 is embedding plus AMCs, Tier 1 is a local LLM acting as a decision model with scoring before output, and Tier 2 is the same LLM with chain-of-thought reasoning tokens for the hard cases. The headline number comes from JevBench's chance-corrected benchmarks: 84.62 points at a 44ms median decision time — Opus 4.6-level judgment without the API bill or the latency. It ships with a Python API, a Pac-Man demo showing it is fast enough for game-engine loops, and it works with any local model (the author tested with EmbeddingGemma-2 and Gemma 4 26B-A4B Q4 QAT). The author, a programmer of 20+ years building this as a hobby to make local AI more accessible, is upfront about the division of labor: the code is mostly vibe-coded, but the architecture and algorithms are human-built. That is exactly the pattern worth studying — the cascade design, the tier cutoffs, the scoring-before-output trick are the human contribution, and the agent did the implementation. If you run anything on local models, the tiered-routing idea ports directly to your own stack.