Machine-learning engineering often focuses on building better models and collecting more data. But sometimes the more important decision is how the problem is represented.

In this talk, I share lessons from a personal project where limited labelled data led me to rethink a conventional computer-vision pipeline. Instead of feeding raw images into a model, I used MediaPipe Hand Landmarker to transform visual input into structured 3D hand landmarks, then normalized the representation to preserve relevant information while removing unnecessary variation.

This architectural choice affects data requirements, model complexity, privacy, and inference latency.

The talk explores a broader engineering principle:

Before reaching for a bigger model or more data, ask whether you can change the representation of the problem.