Reference: Mid-Volume Support Triage Agent

Architectural notes

Reference architecture for a SaaS team auto-triaging inbound support
tickets at mid-volume scale.

Why these picks:
- Sonnet 4.6 as primary for nuanced triage; Haiku 4.5 as fallback for
  simple categorization to keep blended cost down.
- Portkey fronting both so model swaps are config, not code.
- mem0 for per-customer continuity (past issues, account context).
- turbopuffer over Pinecone because per-customer namespaces fit the
  serverless pricing model well at this scale.
- Braintrust for nightly eval against a golden ticket set.
- Helicone for real-time cost visibility for non-engineer stakeholders.

Common gotchas:
- Zendesk MCP can rate-limit on bulk reply writes; queue them.
- mem0 fact extraction needs domain-tuning to avoid noisy memories.