pieta AI, a poem sentiment classifier: built and deployed a full-stack web app (Next.js, React, Vercel, Supabase) that analyzes and classifies a poem's sentiment. fine-tuned a BERT encoder on 6k+ texts for a 4-class task (no impact/pos/neg/mixed) with Hugging Face's Trainer API, then built the evaluation pipeline and custom metrics. reached ~88.6% accuracy and a 0.83 F1-score. also fine-tuned a second model via QLoRA.

fraud detector on an imbalanced dataset: built a 3-layer neural network in PyTorch to detect fraud across ~284,800 transactions, only 0.17% of them fraudulent. engineered custom features with Pandas/NumPy, trained with BCEWithLogitsLoss + Adam, and evaluated with precision-recall curves rather than accuracy, achieving 80% precision and 75% recall on the fraud class.

netflix user activity data analysis: cleaned a multi-csv dataset with missing values, misaligned indices, and duplicate entries using Pandas, then engineered features for a binary classification model predicting user churn.

awards: united states academic decathlon individual competition two-time national champion; john locke global essay finalist (econ category); national latin exam golds & silvers

internships: climate change research @ both nasa's gsfc and the smithsonian institution NESST summer internships

hs highlights: salutatorian, cum laude society junior inductee, admission ambassador, model un club prez