This project aimed to build deep-learning-based action recognition models for emergency response, but when I joined in the later stage, the more urgent issue was no longer the model architecture itself.

I focused on error analysis to evaluate label quality and understand how unstable annotations were affecting performance. Based on that analysis, I argued for additional labeling on unreliable cases, which made it possible to improve model accuracy through more targeted data work.

It was one of my earliest experiences seeing how often data quality becomes the real bottleneck before modeling does.

This project dealt with a recommendation setting where sparse interactions made standard approaches especially difficult to use effectively. I quickly studied graph neural networks and built a GNN-based recommender in PyTorch/PyG, using relational structure to recover signal that existing methods were struggling to capture.

The goal was not deployment, but testing whether graph-based modeling could recover useful signal in a setting where more conventional recommendation approaches were falling short. It was an early example of learning a new method quickly and applying it to a problem with a clear structural bottleneck.