Nova
Shipping a curiosity-forward edu platform for offline schools.
Led the design and research for a learning platform with a locally-run AI tutor to make quality education more accessible for underserved schools in South Sudan. Coordinated between engineers, designers, and international partner organizations.
Nova: offline learning
A device that broadcasts an intranet, giving nearby devices access to an edge-computed AI tutor and resources such as Khan Academy and TED Talks.
No-internet schools in South Sudan and Tanzania lack resources for quality education.
Educators at underfunded schools described unreliable internet, insufficient materials, and difficulty keeping up with changing curriculums. These problems prevented students from getting into secondary school and pursuing higher education.
They had a computer lab room, yet it had no computers and the school had faulty connectivity.
Less than 10% of students go to college
Exam-based culture is cutthroat.
Not enough teachers
One teacher serves 50–60 students, and many teachers are under-qualified.
Outdated material
The government keeps changing curriculums, and teachers have a hard time keeping up.
Unreliable internet
Insufficient government funds and civil war make continuous access impossible.
Students lack textbooks, equipment, and Google Search
“With 50 students in a class, it’s hard to satisfy with the material we have.”
“The government gives new requirements each year, but we lack lab equipment for group work.”
Teachers spent weeks to months making material for new government curriculum distributed each year
“It would be great if we could make new quizzes based on the curriculum faster.”
Little room for curiosity on global topics and hands-on experience
Students borrow teachers’ phones just to Google search, despite poor connectivity. Although D-Link routers were available, electrical outages meant internet access was never continuous.
Product
Existing solutions don’t fuel curiosity
Existing hardware only supported disorganized hyperlinks to content that wasn’t culture-relevant and was often outdated.
Affordable hardware limited LLM context, so we designed AI behavior to handle mid- to long-length conversations
The budget hardware lowered our LLM’s available tokens. Responses became slow after roughly five prompts because the AI parsed the entire chat history.
Reduce friction by avoiding “new chat” every five prompts
Engineers expected users to press “new conversation” every five prompts, but the extra latency meant greater friction. For some students this would be their first experience on a computer, so context and “new chat” might not be understood.
Instead of reading the entire history, check the relevance of prior questions
We implemented an AI relevance-checking system. A smaller model checks past prompts for relevance to the current prompt, then approves the larger model to read only useful context. This lowered computing power and sped up response generation.
Shipped to South Sudan
Shipped to two schools for high school students. We ultimately passed the software to our nonprofit partner without the chance to collect longitudinal data.
Learnings
Ask the dumbest questions
Designers often assume limitations into existence, hindering creative thinking in a startup environment.
Think with the tech, not around it
Learning from engineers creates opportunities to design technical behaviors that influence user experience.
Push for research; nothing is fully gated
Partners on the ground were too busy to conduct student interviews for us, so we set up calls directly within their student network on WhatsApp.