Dan Cowsill
← Selected work

01 / Operate / AI-assisted infrastructure

Infrastructure an AI can help operate.

AI as an operations partner for a self-hosted platform, from routine maintenance to commissioning a refurbished server. Runbooks, recovery plans, and human approval keep changes accountable.

Operations / system map Human oversight
Operational context Runbooks · constraints · retained knowledge
Live services Linux / ZFS / Docker
New foundation Proxmox / encrypted ZFS

Operational and evolving

sanitized with firmware-level verification
7 SSDs
migrated or imported onto the new host
3 workloads
destructive and credential-sensitive actions
Human approval

01 / The problem

Make accumulated operational knowledge reusable

My home infrastructure had grown into a real platform: Linux and ZFS storage, containerized services, internal DNS and certificates, backups, and several dependent workloads. The difficult part was no longer standing up services. It was preserving the reasoning behind safe changes.

I built a durable operating layer for AI-assisted administration. Architecture notes capture current state, memories retain discoveries across sessions, and reusable procedures turn fragile one-off fixes into repeatable work with explicit preflight checks and stopping conditions.

02 / Existing operations

Run the live platform with context intact

The agent can assess and help maintain an Arch Linux, ZFS, and Docker environment over SSH. Its procedures cover kernel and storage-module upgrades, service-by-service container updates, post-reboot health checks, private certificate renewal, media diagnostics, and encrypted off-site backup and recovery.

The runbooks account for the environment's actual failure modes. A container update records recovery information first. Primary DNS is updated last, after proving an alternate resolver so the agent cannot strand its own control path. Reboots wait for a verified storage-module build for the incoming kernel.

03 / Replacement commissioning

Turn refurbished hardware into a secure virtualization host

I used the same operating model to commission a refurbished workstation as a Proxmox virtualization host. Seven SSDs were sanitized, with completion checked against the storage controller's firmware event log, before the machine was rebuilt with mirrored boot storage and an encrypted, redundant workload pool.

The finished host uses automated certificates from a private certificate authority, segmented networking, and an isolated virtual network for sensitive workloads. AI-assisted migration runbooks preserved rollback copies, verified transfers with hashes, prepared guest drivers and networking offline, and defined health checks for each cutover.

04 / Limits and judgment

Keep authority proportional to the blast radius

The system distinguishes routine operations from consequential ones. Read-only diagnosis and reversible service actions can proceed directly. Data deletion, credential rotation, storage changes, and removal of rollback copies require a deliberate human checkpoint, even when the broader procedure is agent-executed.

Secrets are resolved at runtime rather than copied into transcripts, and the documentation records accepted tradeoffs and unfinished hardening work alongside successes. That candour matters: reliable operations depend on knowing what remains manual, what has been verified, and where automation should stop.

Keep exploring

CL650 manual retrieval
All selected work ↗