# Virtual AI Infra Team

A local automation loop that discovers, benchmarks and safely upgrades a machine's own LLM inference setup, keeping a stable OpenAI-compatible API address and rolling back automatically if a candidate optimization fails quality checks.

Virtual AI Infra Team automates the manual work of tracking new speculative-decoding methods, quantization schemes and runtime versions for a locally deployed model: it discovers a candidate optimization, freezes an experiment plan, runs a fair test of the candidate against the current baseline, independently judges quality, speed, memory and error gates, and only then switches the running service to the new configuration, keeping the same API address throughout; a failed candidate is never promoted, and the system keeps or restores the previous version.

It is not a new inference engine or another chat interface; it runs on top of an existing local serving stack and only changes which validated configuration serves each request. A published example shows a generation-throughput improvement of about 107% on an Apple M5 Pro Mac, measured with fixed model commits, deterministic sampling and a four-out-of-four quality gate, with the memory cost disclosed alongside the speed gain, and the raw evidence and a verification script are provided so the numbers can be recomputed.

Requirements are an Apple Silicon Mac, Python 3.11 to 3.13, at least 32 GB of unified memory, and roughly 25 GB of free disk space plus first-run access to Hugging Face. Installation is two shell scripts that create a virtual environment inside the repository without touching system Python.

Website: https://github.com/hsj576/virtual-ai-infra-team