Auto-postmortems.
How Meridian drafts the post-mortem before you open the laptop — and how to teach it your team's voice.
What it does
When a budget event closes, Meridian collects the relevant signals,
diffs against the last known-good window, and drafts a post-mortem in
your team’s standard format. The draft sits in Incidents › Drafts for
review before anything is published.
It’s never sent to a customer without a human pressing the button.
The point of the auto-draft isn’t to replace the post-mortem. It’s to remove the blank page so the human can start with editing rather than remembering.
The lifecycle
Configuring your team’s template
The default template covers Summary / Timeline / Contributing causes /
What went well / Action items. To use your own, drop a markdown file at
templates/postmortem.md in your workspace config repo with {{...}}
placeholders for the system to fill.
What it can fill, and what it can’t
Tuning the model’s voice
Meridian uses a small fine-tuned model that runs inside your tenant. To
adjust its voice, drop reviewed post-mortems into templates/voice/ —
the model retrains nightly against the latest 50 examples.
Voice tuning never sees data outside your workspace. The model weights live in your tenant’s KMS, the training data never leaves your region, and we don’t keep a copy.
Guarantees
- The draft is never auto-published. A human signs every release.
- The model is tenant-isolated. It does not learn across workspaces.
- The data used in drafting is inside your retention window — if you shortened retention to 7 days, the model has 7 days of context.