I built face recognition. For cows. The interesting part isn't the animals, it's that the model is a plug.

I built face recognition. For cows. The interesting part is the architecture underneath, not the animals.

The problem: identify an individual animal from an image, per farm, reliably enough to track it over time. The naive build hard-wires one vendor's model into the whole system and prays that vendor never changes pricing or shuts down.

I did the opposite. The pipeline that enrolls an animal, later identifies it, and gates on image quality talks to the model through an embedding interface. Behind that interface I can run a stub for tests, a hosted recognition service, or a different model per species and modality. The identification logic never changes when the model does. All it ever sees is "turn this image into a vector."

That's the same discipline I use for license-plate reading and document extraction. The applied-AI value was never the model itself. Models are commodities that churn every few months. The value is the system around the model that lets you swap it without a rewrite.

Cows, cars, or documents, the pattern holds. Keep the model at arm's length behind a contract.

Curious how I structure pluggable model backends? Ask below.