NurQuran

A fully local AI tutor that runs 100% in the browser. No API calls. No data leaves the device. It uses Transformers.js v4 with a dedicated static Web Worker to keep inference off the main thread.

The default model is Gemma 4 2B Vision (desktop WebGPU) with Qwen 2.5 1.5B as the multilingual fallback. Gemma 4 supports full multimodal input — images, audio, and video — all preprocessed in the worker before inference.

The AI has semantic memory stored in IndexedDB. Archived chat summaries are embedded locally with multilingual-e5-small and injected as RAG context for relevant follow-up sessions. It also auto-grounds Quran queries by calling a local search tool before responding.

A semantic verse discovery engine built around a 30-emotion catalog. Users describe how they're feeling — “anxious about the future”, “grateful” — and the engine surfaces the most contextually relevant ayahs.

The search runs in three tiers all under 100ms: exact token lookup → binary prefix search with typo tolerance → substring fallback. It runs in parallel with the Quran.com API for proper nouns and Arabic phrases. Results are ranked using a curated semantic boosts system and qualitative match levels (Strong / Moderate / Weak).

The full dataset covers 6,236 ayahs with 20+ semantic fields each: divine message, practical takeaway, user intents, emotional triggers, and life situations. A 23-language native synonym dictionary handles colloquial queries like “أنا حزين” or “coeur brisé”.

A full spaced repetition system for learning the Arabic alphabet, letter combinations, vocabulary, and root word families. Cards are scheduled using Dexie IndexedDB and follow SM-2 spaced repetition logic.

The learning flow covers five phases: letter recognition → pronunciation (with real-time speech grading) → letter combinations (Noorani) → vocabulary (1,000 highest-frequency Quranic words) → root derivation (Sarf). Each phase gates progression behind mastery scores.

Daily flow onboards 3-5 new items based on performance, and the deck automatically top-ups after reviews clear so learners never stall waiting for the next day.

The app is built to function fully offline. The service worker handles shell caching, SPA navigation fallback, and proactive version checking against a live app-version.json — users get OTA updates without hard-refreshing.

Two separate service worker scopes prevent PWA conflicts with OneSignal: the main shell worker at root (/sw.js) and the push worker at /push/onesignal/. A versioned APP_VERSION string busts all worker and model caches predictably across deployments.

Large asset updates (130MB+ embedding files) use atomic OPFS swaps: download to .tmp → verify integrity → bulk-insert to IndexedDB → rename files. The user never ends up in a broken state from a dropped connection.

A high-fidelity sharing engine that turns any ayah into a beautiful social media card. Every card uses “Liquid Glass” glassmorphism with a procedural Islamic geometric pattern uniquely seeded by the Surah and Ayah number.

The pipeline handles two output streams in parallel: a pixel-perfect PNG export (using html-to-image at 1x to avoid iOS Safari OOM crashes) and a normalized plain-text stream for Web Share API and clipboard — stripping all HTML tags before sharing.

Scholarly footnotes render as professional circular badges in the exported image. A global FontWarmer pre-loads critical Quranic calligraphy fonts before any capture to guarantee consistent rendering.