Hi HN, I’m the creator of Satya, an open-source AI learning companion designed for classrooms in Nepal who lack powerful computers. The Problem: Most EdTech assumes high-speed internet. In rural Nepal, 60% of students rely on 10-year-old hardware (think 3rd-gen Intel i3s with 4GB RAM) and have zero or spotty connectivity. They are completely locked out of the AI education revolution. The Solution: Satya runs entirely offline on this legacy hardware. It combines a local SLM (Small Language Model) with RAG to answer questions based on the official designated curriculum. The Stack: Model: Microsoft Phi-1.5 (Quantized to Q4_K_M). I found this to be the sweet spot for running CPU-only inference ~10-12s on an i3. RAG: Local ChromaDB stores embeddings of Grade wise textbooks and teacher notes. Diagrams: I couldn't afford Stable Diffusion, so I built a custom logic-based engine (Regex + Templates) to generate ASCII diagrams for concepts like "Photosynthesis" or "Database Schema" in <5ms. Why not just use a bigger model? This project operates under a strict "4GB RAM / No GPU" constraint. I've prioritized accessibility over having the "SOTA" benchmark scores. If it doesn't run on a dusty school laptop, it doesn't help our students. The project is fully open source. I'd love feedback on further optimizing RAG for low-resource environments or specific SLM recommendations! |