How CERC builds its citizen accountability platforms with AI-assisted coding

Our approach rests on three principles.

1. Write the specification before the code. For each app, we first write a reference document in plain language: the roles (monitors, head teachers, CERC analysts, the public), the data model (a school, a visit, a problem, its follow-up until the fix is verified), the question sets taken from our monitoring guides, the confidentiality rules. This document becomes the contract the coding assistant must follow. It is the most important work, and the work only CERC’s team can do: it puts ten years of citizen monitoring practice into words.

2. Let the assistant build, then check in the field. The coding assistant, running directly on our server, generates the modules, screens and deployment scripts from the specification. Each feature is then tested with real monitors: a form that 15-year-old pupils do not understand is a failed form, however good the code.

3. Reuse the model. Once EduCheck was stable, HealthCheck was derived from it by replacing the school with the health centre and the school questions with questions on access to care. GovCheck applies the same pattern to the budget of a province, city, commune or chiefdom: giving one’s views while it is drawn up, checking how it is spent, visiting the projects it funds. What used to take months now takes weeks.

We do not want to sell a miracle. These are the real difficulties.

Connectivity. AI tools work online. In Uvira or the Ruzizi plain, a work session can be cut off several times an hour. We have learnt to work in short sessions, save often and pick up where we left off.

Security. A coding assistant can, by mistake, leave an access key in the code history or grant a permission that is too broad. Our platforms receive sensitive reports; every release goes through a review of permissions and secrets, and the people who report are never named in published data.

Drift. Without a solid specification, AI produces code that looks right but misses the need. The vaguer the request, the vaguer the result. That is why we invest in the specification above all.

Skills. AI does not remove the need to understand what it does. A manager who cannot read an error log or check a database backup will get stuck. We train our team in these basics.