Generative AI is everywhere in the talk, and rarely in the architecture. Yet, integrated with method, it genuinely accelerates digital transformation — including in constrained contexts such as public administrations and educational platforms in Africa. Here is how I approach it on the systems I build for the African market.
Start with the problem, not the model
The first mistake is picking a model before framing the need. Generative AI is not an end: it is a component. The BMAD method applies as-is — understand the business, model the data, then decide where a language model brings measurable value.
Three families of use come up consistently:
- Process automation — drafting institutional correspondence, summarizing files, classifying inbound documents.
- Assisted generation — code, queries, templates, from structured specifications.
- Augmented retrieval (RAG) — querying an internal document base in natural language, with sourced answers.
RAG: the most useful brick (and the most poorly built)
Retrieval-Augmented Generation connects a model to your own documents: the model "knows" nothing about your domain, it cites your sources. That is the difference between an assistant that invents and an assistant that answers. But its quality depends entirely on data preparation.
A crookedly scanned PDF is not data, it is concrete. Before any RAG, you must structure: extract the text, chunk it cleanly, give it hierarchy.
{
"source": "circular-2026-014.pdf",
"section": "Renewal procedure",
"chunk_id": "c-042",
"text": "Renewal is filed at least 30 days before expiry...",
"metadata": { "service": "immigration", "lang": "en", "page": 3 }
}A base structured in Markdown/JSON versions well, backs up cheaply, and feeds a reliable vector index. It is the most overlooked step — and the one that decides everything.
Security is not optional
Connecting a model to institutional data opens a new attack surface: prompt injection, data leakage, over-trust in a wrong answer. My non-negotiable rules:
- Data isolation — a user only retrieves what they are entitled to; filtering happens before the model, not after.
- No automated irreversible action — the model proposes, the human decides on anything that commits.
- Traceability — every answer cites its sources; every query is logged.
- Graceful degradation — if the AI service goes down, the application keeps working without it.
An asset for the public and education sectors
In the government and university platforms I build, these uses are concrete: improving institutional correspondence systems, personalizing learning paths, automating repetitive administrative tasks. The goal is never to replace the human, but to give time back for what matters.
Conclusion
Generative AI delivers when treated as an engineering component: framed by the need, fed clean data, bounded by security, and degradable without breakage. The rest is just a demo. If you want to integrate these bricks into a platform built to last, let's talk.