Nova
Offline-first scholar assistant for rural students.
Nova is an offline-first assistant built for rural students, keeping study material and tools usable without a reliable connection.
Why did I build it?
AI today is really common, and rural people deserve access to this powerful tool too.
Challenges
Yes, rural cities has a lot of tech constraints, one, is the devices used there, they're not as powerful I'd like, so I need to, again, fall back to edge-ai. I benchmarked the Qwen, Gemma, and Llama families, and I choosed Gemma4:E2B as the brain of the system, because it had a good performance on the tests, less hallucination rate, best overall knowledge and reasoning. But there was still a problem, a huge one: everyone was very dumb in math, since basic math equations like 2x + 2 = 8 made the model struggle, I decided to add a basic math engine plus an intent detector. So, if the intent detector detects math intent in the question, it delegates the task to the math engine, so the model just explains, not solves. If the intent is a factual question, it justs delegates the task to the model to asnwer. Pretty simple workflow.
Learning
It was my first experience with harness engineering, and how I use them daily, and design one was funny.