Mudarris AI Tutor

Learning Arabic as a non-native speaker is notoriously difficult: the language has diglossia (Modern Standard Arabic vs. multiple dialects), complex morphology, and a script that beginners find intimidating. Existing language apps like Duolingo treat Arabic as a low-priority language with shallow content and no understanding of the learner's native tongue. Mudarris wanted to fill this gap with an AI-first approach.

The founding team needed an AI system that could assess spoken Arabic pronunciation in real time, identify specific phoneme errors (distinguishing ح from ه, for instance), and provide corrective feedback — all without requiring a live human tutor. This demanded cutting-edge speech recognition fine-tuned on Arabic phonetics, which no commercial API offered out of the box.

Business model complexity added pressure: the MVP had to be lean enough to raise seed funding, yet demonstrate enough AI sophistication to differentiate from cheaper competitors. We had a tight eight-week timeline from concept to demo-ready product.

We built the core AI tutor on top of OpenAI's Whisper model, fine-tuned with a curated dataset of Arabic speech samples covering MSA and three major dialects. A custom pronunciation scoring engine compared phoneme sequences against reference recordings, returning a confidence score and a specific corrective hint — rendered in the learner's native language using GPT-4o.

The adaptive curriculum engine used a spaced-repetition algorithm (inspired by SM-2) combined with GPT-generated lesson content to ensure each session felt fresh while reinforcing weak vocabulary clusters. A React Native app provided the student-facing interface, while a Next.js admin panel let Mudarris's content team manage vocabulary banks, lesson structures, and learner analytics.

We integrated Arabic TTS (text-to-speech) using a high-quality neural voice synthesiser to give learners accurate pronunciation models at normal and slow speeds. The backend was containerised on Google Cloud Run for cost-efficient autoscaling during the pre-launch period, keeping infrastructure costs low while investor demos were underway.