Operational assurance · Vocational training

A guard that checks the other automations

Every 15 minutes a guard checks that sign-ups, payments and access changes really happened, and speaks up only when something is new, still open after six hours, or resolved.

Swimlane diagram. Every 15 minutes n8n loads the incidents it already has open, fetches recent sign-ups from the member portal and checks Airtable for missing trainees, approved payments with no access change and failed role changes. It classifies each finding as new, ongoing or resolved, upserts it into an incident log on a stable key and sends one digest by email and Telegram only when something new appears, a reminder is due or an issue clears. If a source cannot be read it raises a source incident and does not auto-resolve. A heartbeat goes to an uptime monitor.
How the information moves. Colours show what travels along each line; numbers match the steps below. Select the diagram to open it full size; on a small screen it also scrolls sideways.

The problem

A trainee’s journey passes through several separate automations: the portal sign-up reaches the database, an approved payment opens the course, a completed assessment changes the trainee’s access. Each of those works on its own. When one fails quietly, a paying trainee is left without a record or without access, and the first anyone hears of it is a complaint.

How it works

  1. Start from what is already known. Every 15 minutes the guard loads the incidents it already has open.
  2. Ask the portal. It fetches recent sign-ups from the member portal’s API.
  3. Ask the database. It checks that each sign-up older than 20 minutes and younger than 72 hours exists as a trainee record, then looks for approved payments whose access change has not been recorded after 30 minutes, and for assessments whose role change failed or stalled.
  4. Classify. Each finding gets a stable incident key and is compared with what is open: new, ongoing, due a reminder, or resolved. The result is upserted into the incident log.
  5. Speak only when it matters. One digest goes out by email and Telegram when something new appears, when an open issue reaches six hours since its last alert, or when an issue clears. A heartbeat to an external uptime monitor shows the guard itself is still running.

The engineering that matters

Incidents are written on a deterministic key, so a problem seen in ten consecutive runs is one incident with a history, not ten alerts.

The most important rule is about what the guard does when it cannot see. If a data source cannot be read, the guard raises a source incident and deliberately does not auto-resolve that family of checks. An API outage can never be mistaken for “all clear”.

Time windows matter too. The 20-minute grace period lets the normal automations finish before the guard judges them, and the 72-hour ceiling keeps old, already-handled records out of the comparison.

Have a process like this?

Tell us what it looks like today. We will say whether this approach would work for you.