Optimor is a lightweight Transformer-based AI designed to run entirely on mobile devices (Android). Built by Muhammad Awais at age 17, it aims to demonstrate that advanced AI can be created and deployed on consumer hardware.
- Real Transformer architecture – Multi‑head self‑attention, feed‑forward networks, positional encodings.
- On‑device inference – Runs locally using only Python and NumPy.
- English dictionary vocabulary – 204 words (expandable).
- Conversational interface – Chat with the AI in terminal.
| Component | Status |
|---|---|
| Transformer code | âś… Complete |
| Vocabulary loading | âś… Working |
| Forward pass | âś… Working |
| Training / Backprop | ❌ Needs work |
The model initializes correctly and can generate responses, but the output is currently random dictionary words because the training loop does not yet properly update weights. The transformer architecture is sound; we need help implementing efficient backpropagation (or fixing the numerical gradient approach) to make the model learn from conversation data.
- Python 3.7+ (Termux or PyDroid recommended)
- NumPy
- A terminal with storage access
git clone https://github.com/hyboxteam/Optimor.git
cd Optimor
pip install numpypython optimor_final.pyIf the model has not been trained yet, it will attempt to train (slowly) – but currently the training does not reduce loss. You can still chat with the pre‑loaded knowledge base.
We welcome all skill levels! Here are specific areas where help is needed:
· Implement proper backpropagation – replace the numerical gradient loop with analytical gradients (or fix the current gradient calculation) so the model learns from training data. · Optimize for mobile – reduce memory usage, speed up inference, convert to pure Python where possible. · Expand vocabulary – add more words from words.json or support sub‑word tokenization. · Improve response generation – add sampling temperature, beam search, or better fallback strategies. · Create a simple GUI – integrate with a mobile app or web interface.
- Fork the repository.
- Create a new branch for your feature/fix.
- Commit your changes.
- Open a pull request with a clear description.
For major changes, please open an issue first to discuss.
đź“‚ Project Structure
Optimor/
├── optimor_final.py # Main AI script
├── words.json # English dictionary (source: dwyl/english-words)
└── README.md # This file
To create an open‑source Transformer AI that anyone can run on their phone – and to prove that young developers in Pakistan can build world‑class AI tools.
Muhammad Awais – a 17‑year‑old developer from Pakistan, passionate about AI and open source.
· GitHub: @hyboxteam · Contact: [email protected]
MIT – free to use, modify, and share.
If you'd like to support my development work, feel free to contribute through any of these platforms:
+923437335632
1155138880
0xc26f066E2ec5822a5508Ae0455CCA0A236B620d3
Pakistan Zindabad! 🇵🇰 Let’s make mobile AI a reality together.