ezekiellemana/kazi-agent-adtc

Kazi Agent — Offline autonomous workflow agent for African SMEs, built for ADTC 2026 Laptop LLM Challenge. # Kazi Agent Kazi Agent is an offline autonomous workflow agent for African SMEs and field workers. It is designed to turn local operational records into prioritized follow-up plans, draft messages, summaries, and human-review-required next steps. ## Hackathon context This repository is being prepared for the Africa Deep Tech Challenge 2026 / Laptop LLM Challenge. ## Offline-first promise Kazi Agent is intended to run locally. The project will not use cloud inference, external inference APIs, or hosted model services during runtime. ## Initial repo structure ```text kazi-agent-adtc/ |-- metadata.json |-- download_model.sh |-- REPORT.md |-- README.md |-- .gitignore `-- model/ ``` ## Planned runtime The planned runtime is llama.cpp with a GGUF model file. Planned model placeholder: ```text model/qwen2.5-1.5b-instruct-q4_k_m.gguf ``` ## Model status The model downloader is ready for Qwen2.5-1.5B-Instruct Q4_K_M GGUF. Model weights are excluded from git and must not be committed. ## Run the offline MVP The MVP runs local inference through `llama-completion` from llama.cpp. It requires the local GGUF model downloaded by `download_model.sh` and does not use cloud inference. `examples/shop_records.txt` and `examples/community_collection.txt` contain sample input records. `examples/sample_output.md` and `examples/community_collection_output.md` show sample generated outputs. ```bash export PATH="$HOME/llama.cpp/build/bin:$PATH" ./run_kazi_agent.sh ./run_kazi_agent.sh examples/shop_records.txt ``` ## Profiler The first local GGUF model was tested with `adtc-profiler` in Ubuntu WSL participant mode. This was a local smoke-check, not an official ADTC audit result. Performance summary from the local WSL participant run: ```text tokens_per_second_generation: 31.32 first_token_latency_ms: 3202.8 peak_rss_mb: 1818.32 steady_state_rss_mb: 1737.42 throttled: false accuracy: [] ``` Local accuracy was not run because `lm-eval` is not installed. Participant smoke- …