Autopsy Laps

AI agent that investigates production incidents and auto-generates complete post-mortem reports using real Slack and GitHub data - with human approval before publishing to Notion.

Comment

Inspiration

Every engineer knows the feeling. It's 2am. The incident is finally resolved. You're exhausted. And then your manager asks for the post-mortem. You spend the next two hours piecing together what happened from Slack threads, GitHub commits, and memory - while half asleep. We built Autopsy Labs to eliminate that pain entirely.

What it does

Autopsy Labs is an autonomous AI agent that investigates production incidents and generates complete post-mortem reports automatically.

Given an incident ID, it:

  • Fetches real messages from your Slack #incidents channel via Python
  • Pulls GitHub commits made during the incident window via Python
  • Reasons through all evidence using Anthropic's Sequential Thinking MCP
  • Drafts a complete 6-section post-mortem report
  • Routes it through a human-in-the-loop approval step
  • Publishes the approved report automatically to Notion

How we built it

Built entirely on Airia with the following architecture:

Input -> Python Code Node (fetches Slack + GitHub data) -> AI Model 1 - GPT-4o mini (reasons through evidence, drafts post-mortem) ->Human Approval Node (Human In The Loop checkpoint) -> AI Model 2 - GPT-5 Nano (Notion Publisher) -> Output

The Python node uses the Slack API and GitHub REST API to pull real data before the AI model ever sees the incident. This ensures the post-mortem is grounded in real evidence - not hallucinated from training data.

Tools used:

  • Airia Platform (agent orchestration)
  • Slack API (incident thread fetching)
  • GitHub REST API (commit correlation)
  • Notion MCP (automated publishing)
  • Anthropic Sequential Thinking MCP (step-by-step reasoning)
  • GPT-4o mini + GPT-5 Nano

Challenges we ran into

  • Slack MCP tool calling was unreliable with smaller models - solved by fetching Slack data directly via Python instead
  • Notion API block formatting is extremely strict - required careful prompt engineering to get consistent publishing
  • Getting the AI to reference real GitHub commit SHAs as evidence rather than generating fictional ones required explicit instructions in the system prompt

Accomplishments that we're proud of

  • Real multi-system data ingestion - actual Slack messages and GitHub commits flow into every post-mortem
  • GitHub commit SHA 3b952bf is literally cited as evidence in the generated report
  • Complete end-to-end pipeline: incident in -> post-mortem in Notion out, with human approval in between
  • The output quality is publication-ready with zero manual editing

What we learned

  • Python nodes in Airia are more reliable than MCP tool calling for guaranteed data fetching
  • Smaller models skip tool calls when they think they can answer from training data - explicit data injection solves this
  • Human-in-the-loop is not just a feature - it's what makes enterprise teams actually trust AI output

What's next for Autopsy Labs

  • PagerDuty integration for automatic incident triggering
  • Multi-incident pattern analysis to identify recurring root causes
  • Slack bot interface so engineers never need to leave Slack
  • Integration with Jira to auto-create follow-up tickets from action items
  • Support for multiple repositories and cross-service incident correlation

Built With

Share this project:

Updates