Multi-languageEnglish and Urdu shipped — pipeline ready for more

AI-assisted multi-language pipeline.

Multi-language is not bolted on. The platform ships English and Urdu translations across six interface namespaces today, with an AI-assisted translation pipeline that retranslates only the strings that have changed since the last run. Generated content (lessons, practice, certificates) supports any language the institution operates in; the underlying lesson generator is explicitly multilingual via the learner’s native-language attribute.

Six i18n namespaces

The interface is translated namespace-by-namespace.

  • common

    Shared UI primitives, buttons, status labels, error messages.

  • navigation

    Menu items, breadcrumbs, edition switcher.

  • auth

    Sign-in, sign-up, password reset, MFA, role selection.

  • student

    Lesson surfaces, practice, profile, tutor.

  • faculty

    Teacher tools — assignment review, feedback, gradebook.

  • institution

    Administrator surfaces — rosters, dashboards, reports, licensing.

The translation pipeline

What lets a 50,000-key catalogue update in minutes.

  • Hash-based manifest diff

    Every translation key carries a content hash. When the English source string changes, the hash changes. Only keys with changed hashes retranslate. A 50,000-string catalogue takes minutes to update, not days.

  • Placeholder preservation

    The translation model is instructed verbatim to preserve {{handlebars}} placeholders, <html> tags, and Markdown structure. The generation prompt enforces it; the validator double-checks.

  • Batched throughput

    Eighty keys per batch via gpt-4o-mini. The pipeline is shaped for cost — the right model for the right tier of work.

  • RTL discipline

    Each language carries a direction flag. RTL languages — Urdu, Arabic, Hebrew, Persian — flow correctly in lesson surfaces and tutor chat without per-component conditional logic.

Generated content, generated languages

  • Lesson generation is multilingual

    The lesson generator receives the learner’s native language as an attribute. Lessons can be generated directly in Urdu, Arabic, Swahili, Bengali, Persian, English, Spanish, and any model-supported language. The structure (six steps, fourteen block types) is invariant; only the surface language and culturally relevant examples adapt.

  • Tutor adapts mid-conversation

    The tutor’s system prompt carries the learner’s native language. The tutor will switch language mid-conversation when the learner does — and will revert to the native-language clarity rule (no idioms, no culturally ambiguous expressions) regardless of which language is on screen.

Code is truth
  1. 01

    Six i18n namespaces: common, navigation, auth, student, faculty, institution.

  2. 02

    English (en) and Urdu (ur) shipped today across all namespaces and edition portals.

  3. 03

    AI translation pipeline: hash-based manifest diff identifies changed keys; only those keys are retranslated.

  4. 04

    Translation generation model: gpt-4o-mini at 80 keys per batch, preserving {{placeholders}} and HTML tags verbatim.

  5. 05

    RTL support via per-language direction flag in the language reference table.

  6. 06

    Generation-time multilingual support: lesson and practice prompts accept the learner’s native-language attribute; AI Future Lab deployments cover Urdu, Arabic, Swahili, Bengali, and Persian.

Three regions →All capabilities →