Eight products in the group needed AI features. Paying per token to a US vendor, in XAF, for a Cameroonian business, is a cost structure that does not survive contact with the market.
What I did about it
Everything runs on open-weight models on hardware the business already owns.
Eight SDKs so each product integrates against a typed client rather than raw HTTP.
Positioned as the central AI layer for the other eight products rather than a feature of any one of them.
What I rejected, and why
43 tests against 351 endpoints, and much of the wider surface is scaffolding around a real chat vertical slice. The README says so, and so do I.
Forty-three tests against 351 endpoints, and a large part of the surface is scaffolding around one real vertical slice. The README says so, but I would have shipped fewer endpoints and finished them rather than sketching the whole platform.