African Agentic AI Design Challenge · Governance track

Zabuni — the value-for-money file, prepared in minutes, decided by people

An MCP-based agent for a Kenyan county's Supply Chain Management unit. It benchmarks a proposed tender price against published OCDS awards, profiles the supplier, checks the budget, detects contract splitting, cites a source for every finding, then stops for a named human. It never awards a tender.

What you are looking at: three real runs, recorded unedited from the agent's own audit log, summary and filed report. They use the repo's labelled scripted baseline policy in place of the language model so they reproduce exactly; the two MCP servers, guardrails, self-check, human gate and filing are the real code. Data is a synthetic OCDS fixture with fictional buyers and suppliers. For the open-weights model (Qwen 2.5 via Ollama), run make demo.

1 · Request memo (input)

2 · Agent run — every MCP call, in order

    3 · Filed value-for-money file (output, after human approval)