Yazan Bakhshwin

Built an end-to-end ML pipeline (XGBoost, Random Forest, PyTorch) to predict metallicity of RR Lyrae stars from telescope light curves using ~4,000 spectroscopically labeled stars. Applied the pipeline at scale to produce a catalog of ~135,000 stars and validated against independent star-cluster measurements (~0.15 dex agreement), outperforming Gaia’s published estimates. Diagnosed failure modes such as unreliable input periods and training-label coverage; quantified error impact and built a classifier to flag unreliable inputs. Benchmarked CNNs, hybrid models, and self-supervised pretraining on ~270,000 unlabeled light curves on GPU clusters using identical splits. Delivered posters, a talk, a final report, and co-authored a journal manuscript; continued work by extending the pipeline toward Rubin Observatory LSST DP2 photometry.