Upload study material. Get it explained and animated — in Hausa. Powered by Gemma 4.
# Karatu
**Upload study material. Get it explained and animated — in Hausa.**
Built for the Build with Gemma hackathon, GDG on Campus, Ahmadu Bello University Zaria — 25 July 2026.
**Track 1 — Gemma for Local Language & Literacy.**
*Karatu* is Hausa for "reading" or "study".
---
## The problem
A student at ABU opens a PDF of lecture notes written in dense academic English. They understand
the words individually and the idea not at all. The tools that would help — NotebookLM, ChatGPT
study modes — explain in English, produce walls of text, and need a solid internet connection.
Karatu takes that same PDF and gives back a narrated, animated lesson **in Hausa**, in the
explanation style the student picks.
### Why not just use NotebookLM?
- It does not speak Hausa.
- It produces text and audio, not visual explanation.
- Karatu's model is open-weights, so the same pipeline runs on a laptop with no internet.
---
## How it works
Gemma 4 does **four separate jobs**, not one:
| # | Job | Input | Output |
|---|-----|-------|--------|
| 1 | **Segment** | the whole document | teachable chunks with headings |
| 2 | **Rewrite** | one chunk + learning mode | a re-explanation in that style |
| 3 | **Translate** | that explanation | the same explanation in Hausa |
| 4 | **Script** | the Hausa explanation | animation scenes as structured JSON |
Job 4 is the important one. **Gemma never writes animation code.** It emits JSON against a fixed
vocabulary of six visual primitives, and a hand-written TypeScript renderer maps that JSON onto
Motion Canvas nodes. The model physically cannot break the build, and malformed output degrades to
a plain slide instead of crashing a render.
Full detail: `docs/GEMMA.md`. Architecture rationale: `docs/DECISIONS.md`.
### Learning modes
| Mode | Behaviour |
|---|---|
| **Raw** | Faithful to the source, minimal reinterpretation |
| **Explain Like I'm 5** | Heavy simplification, everyday analogies |
| **Story** | Narrated by a recurring stu …