Cash Platform

Data Infrastructure

* Worked in Ads Measurement Team* Worked on data pipeline

• Worked under John Krumm to build an NLP model that evaluates joke level in each sentence• Used multiple pretrained language models to evaluate score by ensembled them through transformer encoder• Achieved .56 RMSE score

* Helped to bootstrap a privacy compliance start-up* Used Trafilatura to perform web scraping to create around 20,000 HTML dataset. Developed NLP software that determines if a company’s privacy policy is compliant with data privacy law using XLNET model. Achieved similar score on OPP-115 dataset as Polisis paper

* As a senior level capstone, reproduced three metrics that were used to determine contextualization for word embedding model for BERT, GPT2, and ELMOhttps://arxiv.org/pdf/1909.00512.pdf* Adjusted Self Similarity metric by adding variance into account. Adjusted metric showed that BERT is more contextualized than GPT2 by 0.3 z-score while original metric showed GPT2 is more contextualized than BERT by 0.4 z-score* Built a static embedding from Bert using PCA by taking first principal component to represent each word. This performed better than glove static embedding in spam classification using SVC

• Worked on utilizing cascading swaps to use it on Bipartite matching to prove Rota’s Conjecture• Approached Rota’s Conjecture by mapping reduction to higher dimension and mapping back with somerestrictions

Tutored various math classes including discrete mathematics, statistics, calculus series, differential equations, and linear algebra.

Hosted study group and attended every class to help the instructor for pre-calculus and linear algebra.

Helped instructors on grading papers such as calculus and statistics.Gave feedback for answer keys