Fiveangle's homelab and machine learning projects

Developer Dave ("Fiveangle") published a collection of homelab and machine learning projects, including a custom inference server that streams Qwen 3 Mixture-of-Expert model experts directly from Apple Silicon SSD to GPU so large models run without the RAM traditionally required. The projects also include single-command deployment of a high-availability DNS environment on Proxmox using Technitium and keepalived, reporting 0.0s-1.3s observed potential total downtime during container upgrades and failover, plus a utility for running commands across groups of LXC containers, a tool letting an agent in an unprivileged LXC container cut its own network access, and GHES-to-EDAC message routing with EEC error injection on ie31200_edac hardware such as the Dell PowerEdge T110-ii.

With the advent of machine learning capable of actually building stuff, my long-time generalist nature finally has an outlet for the fountain of life improvement ideas I never had time for before. Enjoy -=dave Single-command deployment of a resilient high-availability DNS environment on Proxmox using Technitium + keepalived. Impervious to poorly-behaving clients that won’t failover to query a second DNS server issued to them via DHCP looking at you iOS . Managed container upgrades 0.0s-1.3s observed potential total downtime during upgrade/failover , complete with full testing suite on every container upgrade. No Proxmox cluster required: . Comprehensive utility to execute commands or local scripts en sum inside groups of LXC’s running on a Proxmox server, complete with interactive TUI VMID selection if desired: Allow an agent running within an unprivileged LXC container to safely cut it’s own network access on demand for true network resilience testing : Enable GHES-to-EDAC message routing and EEC error injection/testing on ie31200 edac based hardware such as Dell Poweredge T110-ii to enable rasdaemon ECC error logging and reporting in Debian/Proxmox : Custom inference server highly optimized for running Qwen 3 based Mixture-of-Expert MoE models on macOS Apple Silicon systems, that streams model experts directly from built-in Apple Silicon SSD to GPU in order to run huge models without the RAM required to traditionally do so: