Developed a pipeline using unsupervised machine learning to identify six astrocyte subtypes through spatial clustering of spatial transcriptomic data from the Allen Mouse Brain Atlas, discovered 104 astrocyte-specific genes through differential expression analysis with Bonferroni correction across 10+ million cells.

Enhanced NLI robustness by fine-tuning ELECTRA-small using dataset cartography and contrast sets with distractors, reweighting hard-to-learn examples during training to reduce artifacts. Achieved an 8% accuracy increase on new contrast sets after adding distractors and reweighting training examples to improve model robustness.