Research focuses on process-level methods for reliable ML: dataset-wide attribution analysis, interactive visualization, and model debugging to mitigate spurious cues. Also develops privacy-preserving personalization using federated learning with early aggregation and compressed communication.