Campuspadi SmartNotes Lite

The Problem

The most capable AI study tools are currently gated behind stable internet requirements and API paywalls. For students managing limited data budgets and unreliable campus Wi-Fi, cloud-based summarization and quiz generation are often completely out of reach. The barrier preventing widespread use of AI in education isn't a lack of computing power—it's access economics.

What We Built

Campuspadi SmartNotes Lite is an offline reimplementation of the Smart Notes feature from Campuspadi — a live study platform serving Nigerian university students. The pipeline takes a typed student note and generates a faithful summary and quiz questions, running entirely on a quantized 3B language model via llama.cpp — no internet, no cloud, no API fees.

How We Built It

Benchmarked four candidate models (Gemma 2 2B, Phi-3.5-mini, Qwen2.5 3B, Llama 3.2 3B) against real student notes on faithfulness, summary quality, and structured output consistency
Selected Llama 3.2 3B Instruct quantized to Q4_K_M (~2GB, ~3.4GB peak RAM)
Built the inference pipeline with llama.cpp + OpenBLAS for CPU acceleration
Engineered the prompt to enforce single correct answers, grounded distractors, and strict faithfulness to the source note
Validated on an Intel Core i5, 6.8GB RAM, Ubuntu 22.04 — hardware close to the ADTC Standard Laptop spec

Results

Peak RAM: 3.4 GB (well under the 7 GB ceiling)
Generation: 5.86 tokens/second
Efficiency score (Seff): 50.9/100
No thermal throttling detected

What We Learned

The hardest problem wasn't making a model run offline — it was making it faithful. Small quantized models hallucinate confidently. Phi-3.5-mini invented a correct answer that directly contradicted the source note. The real engineering work was prompt design and model selection, not just quantization.