The Core Idea: Local Model Variety
Most AI apps lock you into a single model or force you to the cloud. Mawj (موج) flips this paradigm.
It's a versatile platform that gives you a curated selection of models optimized for different tasks—all running locally on your iPhone.
- 1Apple Neural Engine Models: Blazing fast, optimized for speed.
- 2Custom GGUF Models: Specialized capabilities, like Qwen for superior Arabic support.
- 3Zero Cloud Dependency: Your data never leaves your device.
The Vision
We demonstrated the power of this platform with an Intelligent Event Planning system.
Whether you say "Lunch with Sarah tomorrow at 2pm" or "Schedule a team sync next week," Mawj understands. It automatically routes your request to the best model for the job, extracts the details, and schedules it on your calendar.
It's not just a chatbot; it's an intelligent agent that respects your privacy.
Architecture: The Multi-Engine Brain
Building a production-ready local AI platform required solving complex engineering challenges.
Intelligent Routing
The MultiEngineRouter dynamically selects the best model.
• English task? → Apple Foundation Model (Speed)• Arabic task? → Qwen GGUF Model (Accuracy)• User override? → Manual selection respected
Performance Optimization
We reduced message processing time from 5s to sub-second latency.
How? A hybrid intent detection system. We use instant heuristics (<1ms) for simple commands and only spin up the LLM for complex reasoning.
Modern Swift Stack
Built with Swift 6 strict concurrency, SwiftUI for the interface, and SwiftData for local persistence. The architecture is modular, testable, and rock-solid.
Privacy in Action
