Independent technical work

Homelab AI

Operational question: how do you treat self-hosted inference like an operable system (desired state, validation evidence, and rollback) when you are the only operator?

Open source (FOSS) · solo · not a commercial product · Solo maintainer / operator

View repositoryDiscuss a roleConnect on LinkedIn

Homelab AI is an open-source (FOSS) Ansible configuration-as-code project for a personal self-hosted AI inference lab on a commissioned Dell Precision 5820 with dual Intel Arc GPUs.

Git is the desired state. Two independent containerized llama.cpp SYCL workers run one per GPU behind an authenticated gateway. HashiCorp Vault is the secret authority. Validation probes produce evidence bundles; upgrades track current, candidate, and previous_known_good versions with rollback runbooks.

Implemented work includes bootstrap, convergence, patching, and upgrade automation; the commissioned dual-GPU profile; authenticated gateway patterns; Vault-managed secrets; scheduled reconciliation; and schema-validated evidence bundles.

This is a personal lab for one operator. It is FOSS and not a commercial product as it stands. A commercial offering may be built on it later; that is not claimed here. It is not a multi-tenant platform, managed service, or customer deployment. A second node profile remains uncommissioned.

Implemented

Evidence you can inspect

Operations material for a one-operator inference lab. These are representative examples of topology, validation, and recovery documentation—not a managed service or multi-tenant deployment.

Architecture and recovery path

Two independent llama.cpp workers behind an authenticated gateway with Vault as secret authority. Validation evidence and an explicit previous_known_good path keep recovery written down.

Scope boundaries