Kazi Agent

• 

  Architecture Diagram

• 

  Profiler summary

• 

  Linux terminal project listing with model info

• 

  Kazi Agent workflow dashboard example

Inspiration

Many African small businesses, field teams, community groups, and local operators manage important work using notebooks, spreadsheets, exported CSV files, SMS records, and offline documents. When internet is unreliable or data is sensitive, cloud AI tools are not always practical.

Kazi Agent was inspired by the need for a useful AI assistant that can run locally on ordinary laptops and help people organize work without sending private records to the cloud.

What it does

Kazi Agent is an offline autonomous workflow agent for African SMEs and field workers. It is designed to turn local operational records into:

• prioritized action plans
• short follow-up message drafts
• risk summaries
• missing-information checks
• next steps that require human approval The agent focuses on practical workflows such as customer follow-ups, community collections, field operations, and small business task planning.

How we are building it

The project is being built for the ADTC 2026 Laptop LLM Challenge with a local-first design.

The submission uses the official ADTC structure with metadata.json, download_model.sh, REPORT.md, and a GGUF model path. The planned runtime is llama.cpp with a small quantized GGUF model suitable for low-resource laptops. The first planned model is Qwen2.5-1.5B-Instruct Q4_K_M because it offers a balance between instruction-following quality, multilingual support, and efficiency for an 8GB RAM laptop target.

Offline and privacy design

Kazi Agent does not depend on cloud inference. The goal is for users to keep their records on their own machine while the agent processes tasks locally. The agent does not automatically send messages or take external actions. Instead, it produces draft outputs and marks actions that require human review.

Why it matters

Many teams across Africa work in low-connectivity environments and still need reliable tools for planning, follow-ups, and local decision support. Kazi Agent is designed for those environments: it keeps data local, works toward low-resource laptop compatibility, and produces human-reviewable outputs instead of taking risky actions automatically.

Current progress

Kazi Agent now includes a runnable offline MVP. The repository includes run_kazi_agent.sh and examples/shop_records.txt, allowing the agent to process local SME records through llama-completion and the local GGUF model.

The MVP produces a brief status summary, prioritized action plan, draft follow-up messages, risks or missing information, and a human approval checklist. This strengthens the project from a profiler-ready submission into a working offline autonomous workflow prototype.

Built With

• autonomous-agents
• gguf
• llama.cpp
• localllm
• offline-ai
• python
• qwen2.5
• ubuntu