Patientkwizera/twigire-ai

Offline, voice-first agricultural advisor in Kinyarwanda — built for the 8GB laptop, no GPU. ADTC 2026 submission. # Twigire AI **An offline, voice-first agricultural advisor in Kinyarwanda, grounded in Rwanda's own extension guidance — built to run on the laptop a farmer's cooperative actually owns.** Submission for the Africa Deep Tech Challenge 2026 — Laptop LLM Challenge, Agriculture domain. ## Overview Twigire AI runs a quantised small language model entirely on-device (8GB RAM, integrated graphics, no discrete GPU) and answers agricultural questions grounded in real Rwanda Agriculture Board (RAB) and MINAGRI extension materials — in Kinyarwanda, by voice. - **No internet required** after setup - **No cloud API costs** - **Voice in, voice out** — removes the literacy barrier - **Grounded, not hallucinated** — RAG over real extension documents, not open-ended generation ## Problem domain Agriculture: crop advisory for smallholder farmers and extension officers, focused initially on coffee, potato, and maize. ## Architecture User speech (Kinyarwanda) → Speech-to-text → Query embedding + retrieval over RAB/MINAGRI corpus → Small quantised LLM (Gemma edge variant, GGUF Q4, llama.cpp) → Grounded response generation → Text-to-speech (Kinyarwanda) → Spoken answer ## Repo structure twigire-ai/ ├── corpus/ # Source extension documents + processed chunks ├── src/ # Application code │ ├── rag_pipeline.py │ ├── voice_io.py │ ├── model_config.py │ └── main.py ├── eval/ # Evaluation question set + scoring scripts ├── report/ # ADTC 2026 report template submission ├── docs/ # Screenshots, architecture diagrams ├── requirements.txt ├── LICENSE └── README.md ## Setup ```bash git clone github.com cd twigire-ai python3 -m venv venv source venv/bin/activate pip install -r requirements.txt ``` ## Status 🚧 Active development for ADTC 2026 (deadline: Aug 25, 2026). ## Team [Your name] — Kigali, Rwanda ## License MIT — see LICENSE ## Setup ```bash git clone github.com cd twigire-ai python3 -m venv venv sou …