AI draft replies for our support inbox (human approves)

We get ~200 support emails/day, 60% are answerable from our help center. Build: pipeline that drafts replies grounded in our help-center content (retrieval, not free-writing), a review queue where agents edit + approve, and analytics on acceptance rate. Never auto-send. Python/FastAPI backend; plug into our existing React agent console via API. Experience with evals/guardrails matters more than model choice.