ADTC 2026 Submission — Corporate/Enterprise Track
Offline AI that turns a student's typed study note into a faithful summary and quiz — no internet, no cloud, no data plan required. Built from Campuspadi, a live edtech platform serving Nigerian university students.
Nigerian university students face a real access barrier: the AI-powered study tools that could help them most — summarization, quiz generation, concept breakdown — require stable internet and API subscriptions they often can't afford or access. A student in a ESUT lecture hall with 200MB of data left shouldn't have to choose between using it for study tools or saving it for something else.
Campuspadi SmartNotes Lite runs a quantized 3B language model entirely on-device — no cloud calls at inference time — to generate:
- Summaries — 3–5 sentence faithful condensations of the student's own note
- Quizzes — 3 multiple-choice questions grounded exclusively in the note's content
This is a scoped, offline reimplementation of the Smart Notes feature already live in Campuspadi, validated against real student notes from ESUT (Enugu State University of Science and Technology).
Designed for the ADTC Standard Laptop profile:
| Component | Spec |
|---|---|
| CPU | Intel Core i5 / AMD Ryzen 5 (x86-64) |
| RAM | 8 GB DDR4 (7 GB usable budget) |
| GPU | Integrated only — no discrete GPU |
| OS | Ubuntu 22.04 LTS |
- Base model: Llama 3.2 3B Instruct
- Format: GGUF Q4_K_M
- Runtime: llama.cpp
- Size: ~2.0 GB
Selected after benchmarking Gemma 2 2B, Phi-3.5-mini, Qwen2.5-3B, and Llama 3.2 3B against real student notes on the summarization + quiz generation task. Llama 3.2 3B produced the best combination of faithfulness, summary quality, and structured output consistency.
# 1. Download the model (idempotent — safe to re-run)
bash download_model.sh
# 2. Install dependencies
pip install -r requirements.txt
# 3. Run the pipeline
python src/cli.py --note sample_notes/computing_systems.txt
# Summary only
python src/cli.py --note sample_notes/photosynthesis.txt --mode summary
# Quiz only
python src/cli.py --note sample_notes/nigerian_economy.txt --mode quiz
# Pass note text directly
python src/cli.py --text "Your note text here"
# JSON output (for programmatic use)
python src/cli.py --note sample_notes/computing_systems.txt --json.
├── src/
│ ├── cli.py # Core pipeline — note in, summary + quiz out
│ └── prompts/
│ └── smartnotes.txt # Prompt template (single source of truth)
├── sample_notes/ # Anonymized student notes for testing
├── bench/
│ └── results/ # Raw benchmark output from llama-bench + adtc-profiler
├── model/ # .gguf file lives here (not committed to git)
├── REPORT.md # Technical writeup
├── metadata.json # ADTC submission schema
├── download_model.sh # Idempotent model downloader
├── requirements.txt # Python dependencies (requests only)
├── AGENTS.md # Agentic AI briefing file
├── SKILLS.md # Recurring task playbooks
├── LESSONS.md # Running log of fixes and learnings
└── VERIFIER.md # Definition of done checklist
Why no RAG? Retrieval over external corpora adds a second engineering surface (embedding model, corpus curation, retrieval quality) that competes for the same 7 GB RAM budget as the LLM itself. The core value — grounding the output in the student's own note — doesn't require retrieval. We deliberately scoped this out and documented it rather than building a half-working retrieval layer.
Why Q4_K_M? The K-quant suffix preserves more weight information in attention layers than naive Q4, giving meaningfully better output quality at the same file size. Q4_K_M is the established practical sweet spot for accuracy vs. RAM on this hardware band.
Why llama.cpp? CPU-only inference with OpenBLAS acceleration, no CUDA dependency, native GGUF support, and direct integration with the ADTC profiler toolchain. No abstraction layers between the model and the hardware.
Measured on Intel Core i5-6300U, 6.8 GB RAM, Ubuntu 22.04 LTS, CPU-only. Raw data in bench/results/.
| Metric | Value |
|---|---|
| Generation speed | 5.86 tokens/sec |
| Peak RAM (RSS) | 3,436 MB (3.4 GB) |
| RAM headroom | 3.6 GB under 7 GB ceiling |
| Thermal throttling | None detected |
| Efficiency score (Seff) | 50.9 / 100 |
| params_match | ✅ Verified |
Campuspadi is a live study platform built for Nigerian university students, offering Smart Notes, AI Tutor, Study Groups, and CGPA tools. SmartNotes Lite is an offline proof-of-concept of its core note intelligence feature, built to demonstrate that the access-economics barrier — not the technology — is the real bottleneck.
ADTC 2026 | Corporate/Enterprise Track