Offline AI voice tutor for blind and underprivileged children in Africa — powered by Gemma 4 (Ollama), Whisper STT, and Kokoro TTS. No internet required.
# EduVoice 🎓🔊
### Offline AI Voice Tutor for High-Performance Learning in Low-Connectivity Regions
EduVoice is a fully offline, voice-first AI tutor designed to bridge the educational divide in regions with limited or no internet access. By transforming consumer-grade hardware (Raspberry Pi/MacBook) into an interactive **"Teacher in a Box,"** it provides students with high-quality, personalized tutoring through a specialized, fine-tuned Gemma 4 model.
---
## 🛠 Technical Implementation
### Unsloth Fine-Tuning
The core intelligence is powered by **Gemma 4 (E4B)**, fine-tuned using the **Unsloth** library. This optimization allowed for a 70% reduction in memory usage and 2x faster training, resulting in highly specialized LoRA weights (`/eduvoice_lora`) that prioritize academic reasoning and pedagogy over general-purpose chat.
### Headless CLI Architecture
To maximize hardware performance on edge devices, the system operates entirely through a **headless CLI loop**. By eliminating the overhead of a web-server and browser interface, an additional 2GB of RAM is reclaimed for LLM inference and high-fidelity vocal synthesis.
### GGUF Quantization Pipeline
The model is deployed as a **4-bit quantized GGUF (q4_k_m)** via Ollama. This specific quantization method ensures the model retains nearly all of its educational logic while remaining small enough to run on devices with only 4GB–16GB of RAM.
---
## 🏗 Tech Stack
| Component | Technology | Role |
| :--- | :--- | :--- |
| **LLM** | Gemma 4 (Fine-tuned) | Core reasoning and tutoring logic. |
| **Inference** | Ollama | Local GGUF serving via Modelfile. |
| **Ear (STT)** | Whisper (Small) | Fast, offline speech-to-text. |
| **Voice (TTS)** | Kokoro TTS | Lightweight, human-like voice synthesis. |
| **RAG** | ChromaDB | Local vector indexing for syllabus grounding. |
---
## 📁 Project Structure
```plaintext
eduvoice/
├── eduvoice_lora/ # Fine-tuned LoRA adapters
├── finetuning/ # Training notebooks and d …