tmachingur-code/adtc-shona-coding-tutor

An offline, low-cost AI coding mentor built for the ADTC Laptop LLM Challenge 2026. Runs a quantized LLM locally via RAG over a curated CS syllabus, with bilingual (English/Shona) explanations — no internet, no API costs, no cloud dependency. # Offline Shona AI Coding Tutor An offline AI coding tutor built for the **Africa Deep Tech Challenge 2026 — Laptop LLM Challenge**. Teaches Python and Computer science fundamentals in English and Shona, running fully on-device with no internet required. ## Features - Fully offline inference (Gemma-2-2b-it, quantized GGUF Q4_K_M, via llama.cpp) - Bilingual: English and Shona, for both explanations and practice questions - RAG-grounded answers from a curated **58-topic** CS syllabus (Python basics through OOP, recursion, lambdas, and debugging strategy) - Two modes: ask a question, or generate practice questions on a topic - Runs within a 7GB RAM budget, CPU-only (official profiler measured peak: **2.75GB**) - Attempts a real answer even for questions outside the syllabus, instead of a flat refusal — see How It Works ## Setup ```bash # 1. Clone and enter the repo git clone github.com cd adtc-tutor # 2. Create virtual environment python3 -m venv adtc-tutor-env source adtc-tutor-env/bin/activate # 3. Install build tools sudo apt update && sudo apt install build-essential cmake -y # 4. Install dependencies pip install -r requirements.txt # 5. Download the GGUF model + MiniLM embedder into model/ (public, no credentials) bash download_model.sh # 6. Build the RAG index python3 build_index.py # 7. Run the tutor python3 rag_tutor.py ``` ## Usage You'll be asked to choose a mode and a language once, up front, and that choice applies consistently across both modes. ``` Choose mode - (1) Ask a question, (2) Get practice questions: 1 Language (english/shona) [1=english, 0=shona]: shona Ask a coding question: Chii chinonzi for loop --- Tutor's Answer --- Musoro: For Loops Tsanangudzo: For loop inodzokorora chikamu chekodhi kamwe nekamwe... ``` ## How It Works Your question is embedded (locally, via a shared `embedder.py` module wrapping `all-MiniLM-L6-v2`, loaded from `model/` with no network calls) and matched …