Bonsai-Llama-Jev: Fast, Vision-Enabled Local Type-Safe AI Decision Engine

A developer released Bonsai-Llama-Jev, a custom llama.cpp fork that enables local 27B vision-language models to serve as TypeSafe AI-compatible System One decision engines using under 10 GB VRAM. It achieves ~76% soft accuracy across a 22,000-request decision benchmark with a median latency of roughly 171 ms. This development allows developers to run high-speed, well-calibrated, type-safe decision making and vision capabilities locally on consumer GPUs without relying on cloud APIs. By unifying traditional OpenAI chat completions with structured System One decision APIs into a single model, it eliminates the need to run multiple specialized models locally. The system introduces a post-hoc model calibration method using a proposed `calibration.json` standard, achieving an Expected Calibration Error (ECE-15) of 13%. It supports multimodal inputs (text, JSON, base64/URL images) through a new `POST /v1/systemone` endpoint while maintaining 100% compatibility with official TypeSafe SDKs and standard OpenAI API endpoints.