Multilingual Email Orchestration

Open the Inbox - the product surface. Each inbound email arrives triaged: its detected language, the classified product and issue, an advisory churn flag and a confidence score. Open one to see the cleaned message, the enriched account facts (from a clearly-labelled simulated backend lookup), and the generated draft sitting beside its QC result - each check passed or failed, with any claim flags called out.

From there the agent acts: Approve and send, Edit, Rewrite with feedback (which re-runs draft and QC), Reject, or Escalate. Sending is simulated and labelled - there is no real-send. A low-confidence case offers top-N suggestions rather than forcing a choice; an un-draftable case falls back to a template starter the agent finishes by hand.

Open "Behind the scenes" to see the full nine-stage trace on the uniform trace-row contract - stage, provenance, model, tokens, cost and confidence - with deterministic and local stages ($0 in both modes) clearly separated from the two metered LLM stages. A Cloud/OSS toggle and a cost meter sit here; the meter only moves on draft and QC, and fails closed to the $0 OSS path at the budget cap. Recorded Cloud-vs-OSS divergences are shown honestly, including one "not captured" case. Everything is labelled: synthetic data, simulated backend and send, Cloud live or OSS recorded.

The Console is one unified, role-gated operator surface - use the View-as switch to see the gating. Overview shows volume, language and issue mix, automation and escalation rates, average confidence and Cloud/OSS cost. A governance section models staged automation: how far the system is trusted to act, with sensitive issue types pinned to manual no matter what. The owner-only Configure section tunes templates, classifier thresholds, the routing roster and the grounding rules, all audited. The project-page walkthrough video narrates the full tour.