Multilingual Email Orchestration

An email arrives in any of the supported languages. The pipeline cleans it - stripping signatures, quoted replies and clutter, and extracting order or account references - then runs a triage pre-filter: spam, auto-replies and out-of-office bounces are short-circuited here, before anything expensive happens. What remains is routed by a local language detector to the right language queue, then classified by product and by specific issue, with an advisory churn flag for at-risk customers.

If the email references an order or account, fact enrichment looks it up in the simulated backend - status, billing, account state. Only then does the metered work begin: the draft stage assembles an approved template plus the enriched facts plus policy, and the model polishes the wording. A hard grounding rule holds throughout: the draft never asserts a fact that wasn't enriched - an unresolved-slot sentence is dropped, never invented.

Every draft goes through QC before an agent sees it: a verification pass claim-checks the draft against the looked-up facts, the policy, the language and the format, with a capped retry if a check fails. The result lands in the agent's queue with its tags, confidence and the per-check QC outcome attached.

From there a human is always in the loop. A confident, clean draft is one Approve away from a (simulated) send; a weaker case offers suggestions; an un-draftable one becomes a template starter; a sensitive one is escalated. Edits and rewrite-with-feedback re-run the metered stages, and every action is audited - which is also how the system's staged-automation trust and its templates improve over time.