Gemini Enterprise Agent Platform AutoML hits 0.946 PR AUC on car-damage classifier

The article documents a hands-on project using Google's AutoML on Gemini Enterprise Agent Platform, the service formerly known as Vertex AI. The author prepared labeled images of damaged car parts from Google-provided sample data, split into training, validation, and test sets via a CSV. After training, the model was deployed to an endpoint and called over HTTPS.

The evaluation showed strong overall performance, but the author highlights that the small test set of 15 images and limited data per label make the metrics indicative rather than definitive. One label, hood, had lower PR AUC than the others. The deployment also ran into a project-specific restriction that required using automatic resources via the CLI instead of the console's dedicated resources.

The main takeaway is that AutoML can lower the barrier to building image classifiers, but data quantity and quality remain the deciding factors. For teams considering this route, the practical test will be whether they can supply enough labeled images per class and manage ongoing inference costs after deployment.