Work Services Products About Talk to Us
All work
Case study

Tier 1 Telecom Operator, MENA

Multilingual AI knowledge base for a Gulf telecoms operator — English, Arabic, and Kurdish (Sorani) — production-grade architecture validated as a proof of concept.

Telecoms Iraq / Gulf Arabic NLP AI Contact Centre
3 Languages: English, Arabic, Kurdish (Sorani)
<200ms Query response time in production-grade POC
4 Access tiers: staff, agents, customers, API
The challenge

Three languages. One knowledge layer. No engineering involvement to maintain it.

This operator's contact centre agents handle thousands of customer queries daily across three languages: English, Arabic, and Kurdish (Sorani). Their existing knowledge base was static documentation. Agents searched manually, results were inconsistent, and Arabic search quality was poor. Most enterprise knowledge tools treat Arabic as a secondary language, producing mismatched results — particularly with dialect variation and transliteration.

Before committing to a full platform build, the operator needed to answer a concrete question: could AI-powered knowledge retrieval work reliably across all three languages, in a production-grade architecture, with the governance controls their content team could actually operate without engineering support?

What we built

A working system — not a mock-up

We designed and delivered a fully functional proof of concept for a multilingual AI knowledge base. The architecture was production-ready from day one. The POC was a working demonstration of the full system — not a mock-up built to impress in a demo room.

The system covers the full intelligence cycle: natural language query understanding in English, Arabic, and Sorani Kurdish; ranked article results with copy-ready agent responses; automatic knowledge gap detection and logging; and live re-indexing triggered by CMS publish events with no engineering involvement.

Four distinct access tiers are built into the architecture: internal staff, contact centre agents, customer-facing, and API/LLM consumers. Each tier receives appropriately scoped content from the same knowledge layer — no duplication, no separate pipelines.

Technical approach

Three decisions that shaped the architecture

Multilingual embeddings as first-class citizens
Rather than translating Arabic content into English for embedding, each article is embedded per locale. Arabic queries retrieve Arabic results — semantically matched, not string-matched. Gemini Embeddings provided multilingual capability across Arabic, English, and Kurdish without custom model training.
Live re-indexing via webhook
When a content editor publishes or updates an article in Strapi, the search server receives a webhook, re-embeds the content, and updates the index in real time. No restart, no delay, no engineering ticket. The system is maintainable by content teams, not just engineers — a deliberate product decision.
Knowledge gap detection as a core feature
Queries returning low-confidence results are automatically flagged and logged in a prioritised gap report. Agents see a knowledge gap indicator; the content team sees an actionable list. Closing a gap is a one-step workflow: write the article in Strapi, publish, and the system validates the improvement immediately.
Outcome

End-to-end multilingual AI retrieval in a working system

The POC demonstrated accurate Arabic and Kurdish query understanding, ranked results under 200ms, live re-indexing, and a knowledge governance workflow the content team could operate without touching code.

The architecture is designed for full production deployment with no structural changes required. The POC validated both the technical approach and the operational model — giving the operator confidence to commit to a full platform build on proven foundations.

Building a knowledge platform for your contact centre?

We've done this in Arabic, English, and Kurdish. We can build it for your languages.

Talk to us