Inside a working integration

The request is easy.
The boundaries matter.

Natural language requests become structured plans, validated actions and recoverable writes across Telegram, Trello and Google Calendar.

I own workflow rules, tool integration and acceptance. Claude and Codex support implementation. Explore three behaviours from the deployed personal service below.

A complete request

Request or event

Schedule the weekly review on 15 October 2026, 14:00 to 14:30, Berlin time.

Plan or integration step

Create event · 15 Oct · 14:00 to 14:30 · Europe/Berlin

Observed check result

Accepted by the service validator

The confirmed date and duration produce 2026-10-15T14:00:00+02:00. External writes are mocked in this example.

The implementation decision

AI interprets the request; explicit validation controls whether an action is allowed.

Synthetic examples with recorded validator and integration test results. This walkthrough makes no model calls and writes no tasks or calendar events.

Implementation checks

9 checks passed on 11 October 2026. Offline tests against the implemented Python service. External Trello and Calendar calls are mocked. These checks do not measure live service uptime.

  • ✓access requires sender and private or allowlisted chat
  • ✓calendar retry uses same id
  • ✓date only is local end of day
  • ✓dst gap and fold rejected
  • ✓event requires end
  • ✓nonexistent day rejected
  • ✓query cannot mutate
  • ✓trello retry recovers existing card
  • ✓vague date not fabricated
Inspect the recorded results JSON →

Taking the same pattern into a team

A team rollout would start with one process owner, a limited set of actions and review of real failures. I would measure completed actions, corrections, duplicate incidents and the time users spend resolving exceptions before expanding the scope.

Discuss an AI role →