Mwalimu Code — Offline AI Programming Tutor

screenshot-en
screenshot-pt

Inspiration

University and bootcamp students across Africa learn to program on low-end laptops with expensive, unreliable internet. Cloud tutors (Copilot, ChatGPT) are blocked by subscription cost and connectivity. When a student in Luanda or Nairobi hits an IndexError at 11 pm with no data bundle, they are stuck. I built Mwalimu Code ("mwalimu" = teacher in Swahili) so the tutor lives on the laptop.

What it does

Mwalimu Code is a fully offline programming tutor: it explains error messages, reviews exercises, teaches concepts and generates practice problems — in English and Portuguese, for lusophone Africa. It is tuned to teach, not just solve: answers explain why the code fails, show a minimal example, and flag common student mistakes.

How we built it

Model: Qwen2.5-Coder-1.5B-Instruct, GGUF Q4_K_M (~1 GB) — coder-specialized small models beat generalist peers at code explanation, and 1.5B leaves huge RAM headroom on the 8 GB profile.
Runtime: llama.cpp (CPU-only). The same llama-server binary serves an OpenAI-compatible API and the zero-dependency web UI (vanilla HTML/CSS/JS, streaming with a live tok/s meter).
Validation: benchmarked with the official ADTC profiler in participant mode, using a SIMD-disabled llama.cpp build that replicates the audit sandbox.

Challenges

Making a 1.5B model feel like a tutor (system-prompt pedagogy engineering), keeping the whole stack offline and dependency-free, and honest benchmarking on below-profile hardware (i5-7300U, 2 cores) with sandbox-parity builds.

Accomplishments

Peak RAM of 1.1 GB — 16% of the 7 GB budget (Seff ≈ 84.6) with zero OOM risk; a real bilingual product, not just a model; fully reproducible setup (download_model.sh + run_tutor.sh).

What we learned

Small coder models + careful pedagogy prompting outperform bigger generalist models for tutoring; sandbox-parity measurement avoids nasty surprises at audit time.