Nordstrom's AI enablement team built an agentic feedback loop called NordAssist that automatically converts Slack support conversations into documentation pull requests. When engineers answer developer questions in Slack, the system uses Temporal workflows to monitor threads, wait for conversations to cool off, extract knowledge gaps via AI classification, perform semantic search against a vector store to find relevant docs, and generate targeted PRs with reviewer assignments. The architecture uses a parent-child workflow pattern for parallel processing and fault isolation. Key lessons from V1 to V2 include replacing AI-based resolution detection with simpler timer logic, replacing URL-dependent doc discovery with vector search and two-step relevancy scoring, and keeping diffs small and focused to improve PR acceptance rates.
Sort: