work

vehicle modeling & software team member

longhorn racing electric · ut austin

vehicle-dynamics modeling and simulation software for the team's electric race car, mostly in python.

• python
• ros 2
• gazebo
• modelica / fmi

research assistant

zhongzheng fu lab · ut southwestern medical center

decoded human cognitive control from intracranial eeg recorded in epilepsy patients during a stroop task: predicted errors ~630 ms before the response, decoded errors and conflict from single trials (auroc 0.81 and 0.77), and read out reaction time from neural activity (spearman ρ 0.83).

built the pipeline end to end in python: neuralynx recordings from 3 patients across 25 sessions, bipolar re-referencing, artifact rejection, and multi-band spectral features over amygdala, hippocampus, mcc, ofc, and pre-sma; leave-one-subject-out cross-validation and benchmarks across linear, tree-based, and deep models. mapped δ/θ conflict signals in mcc and fdr-corrected mcc–pre-sma theta coupling under conflict.

• ieeg
• neural decoding
• python

machine learning research intern

mantis ai · mit csail / kellis lab

implemented transformer-based protein–protein interaction models for drug discovery pipelines in python and pytorch; shipped reviewed prs with unit and integration test coverage into an active research codebase.

• pytorch
• transformers
• drug discovery

undergraduate researcher

azad bioinformatics lab · university of north texas

built an esm-3 attention-mil classifier that predicts bacterial pathogenicity from whole proteomes (auroc 0.961, macro f1 0.900), with embedding pipelines on tacc lonestar6 processing thousands of bacterial genomes.

domain-informed feature engineering lifted macro f1 by 10+ points over naive baselines. first author of a springer nature book chapter on ml-based pathogenicity classification.

• esm-3
• attention-mil
• slurm / tacc