Inspiration
Noor addresses an acute crisis affecting millions of girls. When the Taliban banned girls' education in 2021, our teammate Dunya witnessed her friends and family lose access to schooling. Many don't have access to reliable internet, making learning from home nearly impossible. This drove us to create a solution that works without internet and within current restrictions - an AI tutor that brings education directly to girls who need it most.
What it does
Noor is an offline AI tutor that girls can speak with naturally in both Farsi and English. It provides personalized guidance for college applications, answers questions about educational opportunities, and offers academic support - all without internet access. The conversational interface makes learning feel natural and supportive, similar to talking with a knowledgeable friend who understands their unique circumstances.
How we built it
We fine-tuned Gemma 2B model with context-specific knowledge about education access, college applications, and the unique challenges faced by Afghan girls. Our pipeline includes:
- Speech input using Google Speech Recognition
- Local inference using Fine-tuned Gemma via Ollama
- Coqui text-to-speech model for natural Farsi responses
- Everything runs offline on a Raspberry Pi 5, making it truly accessible anywhere
Challenges we ran into
- Ensuring accurate Farsi language processing throughout the pipeline
- Optimizing the model to run smoothly on limited hardware
- Dealing with hardware constraints (had to do a last minute dongle run when we realized the audio jack is missing on Pi 5)
- Balancing model size with performance for offline use
- Creating culturally appropriate responses that reflect local realities
- Working as a remote team made it more difficult to test the project live
Accomplishments that we're proud of
Created something that could make a real difference in girls' educational access
What we learned
The challenges of multilingual AI development and complexity of building offline-first solutions
What's next for Noor
We absolutely want to make this project a reality. This is a real and urgent problem.
- Hardware optimization: Finding the most cost-effective hardware configuration to make Noor widely accessible
- Enhanced learning materials: Including offline educational content, interactive exercises, and games
- Personalization: Implementing memory features so each Noor device builds context about its user over time
- Expanded subjects: Adding comprehensive tutoring for English, math, science, and history
- Community feedback: Working with Afghan girls to continually improve the system
- Distribution strategy: Developing safe ways to get devices to girls who need them
Thumbnail source: https://www.abc.net.au/news/2023-03-23/afghanistan-taliban-ban-high-school-girls-education-impacts/102107162
Built With
- gemma
- python
- raspberry-pi


Log in or sign up for Devpost to join the conversation.