afdroiddev-oss/smart-tutor-adtc-2026

Offline-first CPU maths & science tutor for African learners with disabilities — Africa Deep Tech Challenge 2026 (math_scientific_reasoning) # SMART TUTOR **An offline-first mathematics and science tutor for African learners, running entirely on a CPU laptop.** Submission to the **Africa Deep Tech Challenge 2026 — The Laptop LLM Challenge**, domain `math_scientific_reasoning`. No internet. No GPU. No API key. A GGUF file, llama.cpp, and 8 GB of RAM. --- ## Why Mathematics and science teaching in much of Africa runs into the same wall from several directions at once: intermittent connectivity, intermittent power, laptops that are five to ten years old, and almost nothing built for learners with disabilities. A cloud tutor solves none of those. A tutor that fits in 1.1 GB of RAM, answers with no network, and renders its answers for a screen reader solves all four. --- ## Quick start ```bash git clone && cd smart-tutor-adtc-2026 bash scripts/setup.sh # python env + llama.cpp (prebuilt or from source) bash download_model.sh # fetch the GGUF weights (~940 MB) export PATH="$PWD/vendor/llama.cpp:$PATH" smart-tutor ask "Solve the equation 3x + 7 = 22" ``` Expected output: ```text To solve the equation 3x + 7 = 22, we follow these steps: 1. Subtract 7 from both sides: 3x = 15 2. Divide both sides by 3: x = 5 3. Check: 3(5) + 7 = 15 + 7 = 22 ✓ FINAL ANSWER: 5 ``` --- ## Measured performance **SMART TUTOR** (fine-tuned) Q4_K_M, `llama-bench -p 512 -n 128 -ngl 0 -t 6` (the exact invocation the ADTC profiler uses), 5 repetitions, on an AMD Ryzen 5 5600U: | Metric | Measured | ADTC component | |---|---|---| | Generation throughput | **17.11 ± 0.44 tok/s** | `S_perf` **100.0** | | Prompt processing | 83.28 tok/s (mean) | — | | Peak RSS | **1700.9 MB** | `S_eff` **76.27** | | Steady-state RSS | 1654.9 MB | — | | Thermal throttling | none observed | no penalty | | Held-out accuracy | **29/30 (96.7%)** | — | | Model file | 940.4 MB, 1.54 B params | — | The ADTC profiler's own independent measurement of the same artifact agrees closely: 17.39 tok/s, 1706.9 MB peak RSS, `"measure …