A camera at a gate reads your license plate and decides whether the gate opens. I built that whole loop without sending a single image to the cloud.

Two constraints made it interesting.

Privacy first. Residents don't want their car's movements streamed to some third-party API all day. So plate recognition runs on a self-hosted open-source model, and the raw image bytes never leave the property. The system stores the decision it made, not a running feed of who came and went.

Then swappability. The recognition model sits behind a clean provider interface. The access-control logic asks "what plate is this?" and doesn't care whether the answer comes from an open-source reader, a paid API, or some future model I haven't picked yet. I swapped a commercial vendor out for the OSS one without touching the logic that grants access.

That's the pattern for applied computer vision that survives the real world: keep the model behind an interface, and keep the sensitive data on-site.

If you're putting vision into a product and worried about privacy or vendor lock-in, ask me how I structured this. It's the part people underestimate.