Bhavya Shah Email & Phone Number

Menlo Park, Ca, Us

Washington, Dc, Us

Palo Alto, Ca, Us

I was part of a Scrum-based team building a document AI product that improves layout analysis. I trained YOLOV5 on both programmatically and manually annotated data sets (PubLayNet, DocLayNet, DocBank, E^2, etc.). I developed a working understanding of YOLOV5 hyperparameters and used hyperparameter evolution and manual adjustments to optimize results. I researched different training evaluation metrics, finding that mean average precision and F1-score were good bets, implementing mAP@[.5:.95]. I coded data set generation in Python, using PDFMiner to extract semantic structure from PDFs and pitting those against well-formatted ePubs from the Bookshare library to generate training data. I used EC2 instances from Amazon Web Services for cloud compute, submitted pull requests using Git, and documented issues and contributions through Jira tickets and stories. By the end of the summer, we had a trained model that achieved MAP scores approaching .6 integrated in the pipeline.

Research assistant on the Candidate Priming project for Dr. Jon Krosnick (Stanford) and Dr. Matt Berent (independent consultant). I started off by reading several seminal publications on priming to teach myself about issue sailiance, source credibility, and campaign effects. I conducted a rigorous literature search on these topics using Google Scholar, JSTOR, and Searchworks, and collected PDFs of relevant articles. Then, I co-authored the pre-registration paper for our study, which exposed me to different types of experimental variables (dependent, independent, control, intervening, and extraneous) and to coding a questionnaire (random sampling, attention check, control condition, and order effects). Finally, I self-learnt Stata - reading from a data set, computing a variable, all the way to running linear and logistic regressions with nearly 50 predictors - in order to test 4 hypotheses using data of 6000 responses across 2 surveys. Additionally, I coded the logistic regressions for a co-worker for an unrelated study in R.

Co-organizer of the inaugural Equittable Design in Tech conference at Stanford University. I moderated a panel on the intersection of disability, design, and AI featuring leaders from the US Access Board, Algorithmic Justice League, and academia. Additionally, I was part of the Speaker/Workshop and Marketing/Outreach subcommittees. As part of the former, I identified and sent speaker invites to prospective presenters, including executives at Google and Netflix. As part of the latter, I helped solidify the conference's brand, including its name, theme, structure, and itinerary.

Research assistant on the Campus Human Rights Index project for Dr. Kiyoteru Tsutsui (Stanford), Dr. Charles Krabtree (Dartmouth), and Dr. Volha Chykina (URichmond). I automated the collection of data for 150,000+ courses offered at various US universities by self-learning AutoHotKey and independently developing, optimizing, and executing a script that scrapes webpages using keyboard shortcuts, thereby completing a task allocated 3 months in under 3 weeks. I helped prioritize disability rights as a factor for the index’s rankings and identified 7 innovative manifest variables to measure universities’ commitment to it. I collected data for one of these variables and offered preliminary analysis on the the others.

Us

I Contributed ideas related to product development, financial services, and student engagement through monthly consultations with top executives of the SFCU (including the CEO and CIO). In particular, I emphasized the increasing interest in kryptocurrency trading, suggested utilizing certain platforms to enable cheap international payments, and synthesized the student perspective on customer experience during the pandemic. As a result, the credit union switched to Zelle for payment processing and integrated bitcoin trading services.

Stanford, California, Us

Research assistant on the Policing and Violence Against Women in India project for Dr. Nirvikar Jassal. I encoded the genders of 40,000 complainants in a novel data set of crime records. I quickly developed cultural familiarity with Punjabi and Islamic names to better predict genders, and used bookmarks in Notepad++ and filters in Microsoft Excel to efficiently enrich and clean up the .csv data. Additionally, I explored using NLP to predict genders of Indian names by finding a public data set, cleaning it up, identifying distinguishing features of male and female Indian names (n-gram analysis, ending vowel, initial letter), and implemented an SVM classifier in Python for training.

Mumbai, Mh, In

Produced 19 modules about college readiness aimed at middle school and high school students in India. Each module consists of an approximately 1500-word script for a video, a quiz containing 10 multiple-choice questions, and 3 reflective essay prompts. I covered topics ranging from communication skills and time management to extra-curricular activities and career exploration.

Master Of Science - Ms, Computer Science (Artificial Intelligence Track)

Bachelor'S, Mathematical And Computational Science

High School, Grades 11 And 12, Science Stream (Physics, Chemistry, Mathematics)

School, Kindergarten-Grade 10