Hey folks, are standard face recognition APIs actually accurate enough for real-world datasets, or did you end up building something custom?

We wasted weeks trying to tune standard APIs before realizing they just can't handle weird angles and lighting properly in production. For custom face recognition we didn't want an off-the-shelf API. Went with their site for a tailored model and the accuracy on our dataset was noticeably better. It actually fit our specific pipeline without requiring endless workarounds, which was a huge relief for the team after dealing with so many false positives.