Zakhir.

+ Project overview

Project overview

A bilingual AI support layer for a digital wallet serving thousands of users - handling inbound across WhatsApp, Instagram, Messenger, and web chat in Arabic and English. The Arabic side is powered by a custom Sudanese-dialect model we fine-tuned on open weights and self-hosted, so replies sound native rather than stiff formal Arabic. It handles 87% of inbound conversations end-to-end, escalating the complex cases to human agents.

The problem

1. 01Off-the-shelf Arabic AI defaults to formal Modern Standard Arabic and sounds foreign to Sudanese users
2. 02Support tickets flooding across disconnected channels - WhatsApp, social media, web chat - with no unified view
3. 03Agents overwhelmed by repetitive queries in both Arabic and English
4. 04No after-hours coverage - customers left waiting outside business hours

Our solution

Generic Arabic AI answers in stiff, formal Modern Standard Arabic that reads as foreign to Sudanese customers - so we built a dialect model instead. We fine-tuned an open-weight LLM on a purpose-built Sudanese-Arabic dataset and self-hosted it, keeping the model on infrastructure we control rather than a third-party API. On top of it we built a unified AI support layer across WhatsApp, Instagram DMs, Facebook Messenger, and web chat, with RAG-powered knowledge retrieval, a secure OTP delivery API for the mobile app, automated issue classification, and smart escalation to human agents for the complex cases. The system is hardened against prompt injection and manipulation attempts.

What we delivered

• Custom Sudanese-Arabic dialect model, fine-tuned on open weights and self-hosted
• Unified omnichannel inbox across WhatsApp, Instagram, Messenger, and web chat
• Bilingual support (Arabic + English) with automatic language detection
• RAG knowledge base covering account, payments, and technical support
• Smart escalation with human-in-the-loop handoff
• Secure OTP delivery API integrated with the mobile app backend
• Automated issue classification and visual performance reporting
• Prompt-security hardening against injection, extraction, and manipulation attempts

ROI & impact

Of inbound conversations handled end-to-end by AI

87%

AI responses in production

42,000+

Support sessions handled

8,500+

Two languages, zero night-shift hires

24/7