I'm Sem — I build and operate automation systems in n8n. Anyone can post a screenshot of a workflow canvas. The numbers below are read straight out of the execution log of my own production instance every ten minutes, so they are as current as the system is. If it ever breaks, this page shows that too.
| Date | Executions |
|---|
Scrapes and enriches business listings, generates a personalised outreach draft through the Anthropic API, and posts every draft as an interactive card in Slack with approve, regenerate and reject buttons. Nothing is ever sent without a human click. Separate workflows detect replies and bounces, schedule follow-ups, and post a weekly funnel summary.
A validation layer checks every AI-generated claim against the source record before the draft is shown. An earlier version invented company founding dates; that check removed the failure class entirely rather than patching one symptom.
Everything here runs on a single VPS I administer myself: Docker behind Traefik, managed through Coolify, with Authelia in front of the internal tools. Offsite backups run to Backblaze B2 on a timer — and a scheduled job restores from that repository and verifies the contents, because a backup you have never restored is not a backup. That test caught three defects in my own backup chain that would have made recovery impossible.
The whole server is reproducible from one bootstrap script, with the runbooks in git.
I ask what a workflow is supposed to achieve before I touch a node, and I test against real data rather than happy-path samples. If something I built breaks, I tell you — I don't report "done" on work I haven't verified end to end.
Most of my n8n time has gone into the failures that don't announce themselves: item referencing in multi-item nodes, webhooks registered through the API that silently 404, batching that quietly drops records. Those are the ones that pass a demo and break a month later.