Deeper dive
Weekly eval harnesses compared draft quality against a frozen golden ticket set—no silent regressions when prompts moved.
Grounded answers from tickets, KB, and runbooks — agents escalate with evidence packs, not guesses.
Agents rewrote the same troubleshooting steps daily; tribal knowledge lived in Slack threads outside the ticketing system.
LangChain pipelines retrieve approved articles and prior solved tickets, compose draft replies with citations, and route uncertain cases through a structured escalation chain.
Vector index over curated KB; deterministic tools for billing lookups; supervisor pattern for chained checks before send. No auto-close without human approval.
Normalize payloads from Zendesk plus attachment text extraction.
Hybrid search with recency boosts for policy changes.
Structured JSON output with reasoning traces for QA sampling.
One-click approve/edit; rejected drafts feed few-shot refinement.
Customers saw consistent answers during outages. Coaches finally had grounded examples for training instead of screenshot archaeology.
Weekly eval harnesses compared draft quality against a frozen golden ticket set—no silent regressions when prompts moved.