Turn one Windows box into a private AI server your whole team can use - no installs, no terminals, no data offsite. Per-user API keys, RPM/TPM quotas, a concurrency cap set at the measured peak (8 concurrent on Ryzen AI MAX+ 395, Radeon 8060S, 128 GB), self-healing watchdog, 47 KB single-file admin console, MIT, zero runtime deps, 163 tests. Built for a 20-person firm whose documents legally couldn't leave the network.
