Field notes on what works, what costs money and what fails silently in AI projects across Central Africa.
The problem is almost never the model. It's the choice of use case and the lack of a business owner.
French, English, local languages: what multilingualism really changes in an assistant's architecture.
One page of spec avoids six weeks of rework. Here's the format we use on every engagement.
Inference, monitoring, retraining, on-call: the budget nobody shows in pre-sales.
Traceability, explainability, data retention: three requirements to handle at scoping.
The talent pool exists. What's missing are projects ambitious enough to keep it.