Selected work
A few local AI and self-hosting setups: dsh in containers, a Traefik cutover, a Harbor migration. Client and machine details are anonymised.
What these are
Three projects I can describe without naming the client. Machine details are rounded, but the problems and the fixes are the ones that happened. Code excerpts are simplified; hostnames and paths are placeholders.
1. A workstation that came back after every reboot
In plain terms: the machine now starts itself, and a bad upgrade can be undone.
The problem: dsh was running by hand: one terminal, a model endpoint in a shell profile, plugins installed globally. After a reboot it stayed down, and one upgrade had already broken the plugin set.
What I did: moved dsh into a container with a pinned version, mounted $DSH_HOME and the workspace as volumes, put the web port behind Traefik, injected model credentials at runtime, and wrote down one upgrade and rollback path.
# compose excerpt — pin the image, keep state on a volume
services:
dsh:
image: local/dsh:0.2.2
command: ["dsh", "web", "--port", "8080"]
ports:
- "127.0.0.1:8080:8080" # behind Traefik or cloudflared, not the internet
volumes:
- ./dsh-home:/data/dsh
- ./workspace:/workspace
environment:
DSH_HOME: /data/dsh
restart: unless-stopped
Left behind: the container starts on boot, a backup of $DSH_HOME, a short upgrade runbook, and a verification list (web UI, plugins, sessions).
2. Eight services, one entrance
In plain terms: eight services sit behind one entrance, and certificates renew without anyone remembering them.
The problem: eight self-hosted services on a NAS, each with a port written in a note somewhere, certificates renewed by hand.
What I did: a route table by hostname, Traefik static and dynamic configuration, DNS-01 certificates, and a hostname-by-hostname cutover with the old proxy config kept for rollback.
# dynamic config excerpt
http:
routers:
grafana:
rule: "Host(`grafana.home.arpa`)"
entryPoints: [websecure]
service: grafana
tls:
certResolver: le
services:
grafana:
loadBalancer:
servers:
- url: "http://10.0.20.5:3000"
Left behind: one entrance, automatic certificates, a route table the owner can edit, and the debugging order for 502 and 404.
3. Images the whole team could pull
In plain terms: everyone pulls the same image, and old ones get cleaned up.
The problem: images were built on one laptop and copied around. Teammates pulled whatever tag happened to be on the machine.
What I did: Harbor with two projects and role-based access, storage on the existing array, retention and garbage collection, and a migration from local images.
# harbor.yml excerpt
hostname: harbor.home.arpa
data_volume: /srv/harbor
storage_service:
filesystem:
rootdirectory: /storage
Left behind: versioned pulls, a cleanup routine, backup steps for configuration and database, and upgrade notes.
How I work
- Scope gets written down before work starts, then a quote.
- I take a small number at a time; there is no agency pipeline.
- Configuration, documentation and a way back are part of the delivery.
- Client names and machine details stay out of these notes.
- The full project process, with a sample log.
Next
Not sure where to start?
Run the assessment first, or send your setup and goal by email. The reply says whether it fits and what happens next.