Project: release reviewed localized support replies

Last updated: 6 Oct 20265 min read

project

Translate a support reply only after placeholders, amounts, terminology, meaning and locale-specific review have passed explicit gates.

Define the customer-visible contract

The support agent drafts a reply about a pending refund. The translation must preserve pending status, amount, date, order reference and template placeholders. The agent remains accountable for sending it. Store source revision, target locale, translation bundle and reviewer decision. Unsupported locales return a manual translation request, not a nearest-language guess.

Build the test set

Collect reviewed source replies with repeated placeholders, mixed scripts, product names, negative statements and ambiguous pronouns. Preserve grouping by case and split by time. For each locale, ask bilingual reviewers to mark adequacy, terminology, harmful meaning changes and whether a customer could act incorrectly. Include messages where the source itself needs clarification. The field contract defines deterministic checks.

Choose the release path

Compare the current and candidate translation bundles on the same audit set. Run protected-field checks on every output and bilingual review on sampled high-risk messages. Block release for missing placeholders, reversed refund status or unsupported locale claims. Use locale-slice evaluation to keep a strong majority language from hiding a weak minority locale. Keep the old bundle available for rollback.

Operate human signoff

Return a draft and validation report to the agent. A reviewer can edit or reject it; store corrections under access control and include their reason in the next evaluation cycle. Never send the raw machine output automatically merely because it passed a token check. Monitor correction rate, customer misunderstandings, p95 latency and protected-field failures. Retain only the text and review data needed under the support retention policy.

Implementation

def prepare_localized_reply(source, translated, locale, supported_locales, approved):
    if locale not in supported_locales:
        return {"state": "manual-translation", "reason": "unsupported-locale"}
    import re
    from collections import Counter
    marker = re.compile(r"\{[a-z][a-z0-9_]*\}")
    if Counter(marker.findall(source)) != Counter(marker.findall(translated)):
        return {"state": "blocked", "reason": "placeholder-mismatch"}
    if not approved:
        return {"state": "needs-agent-signoff", "text": translated}
    return {"state": "approved-draft", "text": translated}

reply = prepare_localized_reply("Refund {amount} is pending.",
                                "Refund {amount} is pending.", "hi-IN", {"hi-IN"}, False)
assert reply["state"] == "needs-agent-signoff"

Performance and operating cost

The local marker comparison is O(n + m) time for source and target lengths and O(p) space for distinct placeholders. Translation inference, bilingual review and agent editing dominate total cost. Batch drafts only where privacy and latency allow it; never batch away signoff for high-impact replies. Track both automatic validation failures and meaning corrections so a low marker-error rate cannot mask poor translations.

Common Mistakes

Sending an approved draft without the agent’s final action.
Falling back to a related locale without reviewed evidence.
Ignoring a changed amount because placeholders still match.
Training on agent corrections before checking their source revision and consent policy.