Grounded, Safe, Culturally Aligned, and Scalable AI for Education
Building a $22M AI Pipeline by creating an enterprise-grade AI Orchestration layer.
It all started when...
The Middle East education market faced a paradox: Ministries wanted innovation to achieve the mandates of their national strategies (Vision 2030 in KSA/Egypt, Vision 2040 Oman, Vision 2031 UAE). They were hungry for AI implementations, but they were terrified of cultural misalignment and data sovereignty. While generic tools (ChatGPT, Gemini) existed and were offered for free, ministries and users refused to scale them. Simultaneously, nations like UAE and KSA had invested in local national models (Jaiss, Allam) but these models sat unutilized because they lacked an application layer to bring them safely into the classroom.
My Role
The Digitization Engine: I designed a multi-agent system capable of ingesting 'messy' legacy textbooks (complex PDFs) and transforming them into dynamic learning paths. This saved Ministries millions in manual content digitization costs.
The Ingestion Pipeline: I defined the strategy for a multimodal ingestion engine. Long before it was industry standard, we built pipelines that used vision models to deconstruct legacy textbooks (complex PDFs with diagrams), preserving semantic structure to enable true multimodal retrieval.
The Retrieval Architecture: I co-architected an Agentic Hierarchical RAG system. We implemented multi-hop retrieval to connect related concepts across different chapters, ensuring the AI has the full context before generating an answer.
The Response Pipeline: I designed a dedicated response layer fortified with guardrails and safety checks. This acts as a deterministic filter post-generation to guarantee that no hallucinated or culturally inappropriate content reaches the student.
The Model Selector: I defined a model selector that capitalizes on national investments. The system enables routing to local models (Jaiss, Allam, or on-prem DeepSeek) while reserving expensive cloud models only for complex reasoning.
The Evaluation Strategy: I established a rigorous Evals Framework to benchmark our outputs, proving a drastic reduction in hallucinations compared to base models and building trust with technical stakeholders.
Observability & Cost: I defined the Observability and Cost Management layer. This provides granular audit logs for every student interaction and actively manages token economics, making national-scale deployment financially predictable.
Results
$22M Sales Pipeline generated by proving we could operationalize their National AI assets.
National Pilots secured with Saudi MOE entities, validating the hybrid local/cloud approach.
80% Cost Reduction achieved by routing standard queries to cheaper, local open-weight models.