amabingsarl-debug/urasis-agro-adtc-2026

Offline Kiswahili-first agricultural intelligence for the Africa Deep Tech Challenge 2026. # Urasis Agro — ADTC 2026 Urasis Agro is a Swahili-first, offline agricultural assistant for African smallholder farmers and extension workers. This competition edition uses a 3B-parameter Qwen model in GGUF Q4_K_M format and runs only through `llama.cpp`. The reproducible build pins `llama.cpp` release `b10218` and compiles it with network-backed model loading disabled (`LLAMA_CURL=OFF`). ## Why it exists Cloud assistants require connectivity, recurring API fees and data transfer. Urasis Agro keeps the core interaction on an ordinary 8 GB laptop and can answer in Kiswahili even when the network is unavailable. Its local agricultural notes encourage field observation, low-cost action, integrated pest management and escalation to a local extension officer when uncertainty is high. ## Quick start on Ubuntu 22.04 ```bash cd adtc-2026 bash scripts/setup_ubuntu.sh bash download_model.sh bash scripts/run_cli.sh ``` Run the optional local agricultural retrieval demonstrator: ```bash bash scripts/run_agro_rag.sh --show-sources \ "Majani ya mahindi yangu yana mashimo na kinyesi ndani ya moyo. Nifanye nini?" ``` Run the official profiler: ```bash bash scripts/profile.sh --skip-accuracy # smoke test bash scripts/profile.sh # final participant report ``` Validate the repository structure before profiling: ```bash node scripts/validate_submission.mjs python3 -m unittest discover -s tests -v node scripts/validate_submission.mjs --final # fails while personal metadata is pending ``` ## Offline boundary `download_model.sh` and the one-time setup require Internet access. After the GGUF file and tools are installed, `run_cli.sh` and `run_agro_rag.sh` make no network requests. Gemini, Firebase and WebLLM are not part of this submission runtime. For a strict offline check, disconnect the machine after setup, then run both declared prompts through `scripts/run_cli.sh` or `scripts/run_agro_rag.sh` and the official profiler. The model URL, expected size …