Github Profile

Engineering Experience

Thunkable (YC W16) , San Francisco, CA

Backend Software Engineer

Thunkable is a drag-and-drop programming solution that aims to bring simple design and code techniques to everyone, the platform where anyone can build their own cross-platform mobile apps 

• Reported to VP of Engineering and head of infrastructure 
• Designed and built a client solution to enable user migration from deprecated Android platform to new iOS/Android cross-platform
• Designed and built user-facing features 
• Monitor client reported issues and documentation for new components
• Established a new peer knowledge base download series for interns to share their work, learn from one another, and form community

TeachFX, Palo Alto, CA

Fullstack Software Engineer

TeachFX’s mission is to promote more meaningful and equitable classroom dialogue by superpowering teachers’ work — using technology to provide educators with regular, automated feedback on their practice.

• 3rd engineer on staff reporting to Head of Product and CTO 
• Built new user signup interface that allowed us to group customers by school and identify a new business opportunity the company pivoted to 
• Analytics dashboard (logistic regression), main product feature for clients 
• Built API of nationwide school data that saved lean startup $5K+/year
• NLP work in detecting teacher v. student voice 
• Python scripts to extract user info and automate JSON files conversion +  designed our Firebase user fields structure
• Added Google Analytics to capture user behavior throughout the web app

AI RESEARCH

A Deep Learning Approach for Human Activity
Recognition

The hardware and sensors in smartphones and wearable devices are becoming more powerful and precise every year; the valuable data these sensors collect can be used for more precise Human Activity Recognition. We implemented both Convolutional Neural Networks and LSTMs to classify 18 unique activities ranging from eating chips to dribbling a ball from labelled, time-series gyroscope and accelerometer data collected from a smartwatch and smartphone. All of our work is located in our public repo here, and the processed data can be accessed here.

Prediction of Future Offensive Performance of MLB Position Players

Major League Baseball (MLB) is comprised of thirty professional teams, each of which is valued between $1 billion and $4.6 billion USD. The annual revenue for any given team is in the hundreds of millions of dollars, around half of which is paid to players in the form of salaries. The minimum annual salary in 2019 was $550,000, and the top players earn well over $30 million per year. Because of the market value of MLB players, a team’s decision over which players to sign to a contract and which to let go is crucial to the success of the team both athletically and financially. Unfortunately for teams, the prediction of a player’s future value is difficult, and there are many instances of players earning much more than their performance warrants. We set out to predict the value of an MLB position player–a player who does not pitch–in the next year given the player’s performance over the previous n years. We trained multiple models on data from the 2005 – 2015 seasons and validating on data from 2016 – 2018. We also built individualized LSTM models structuring it as a time series forecasting problem. Paper here.

NLP Hate Speech Detection

Improving Hate Speech Classification on Twitter Presented at the Latinx In AI Workshop 2019 NeurIPS conference 

• Improving Hate Speech Classification on Twitter using Flair, stacked word embeddings, hand built features; 5-fold cross-validation grid search on a Logistic Regression model using hand-built features  and sentence-level embeddings 
• Detailed data analysis resulted in detection of biases in model which downgraded hate speech to offensive when language was directed at women / people of color
• Our model outperformed the baseline model by 2% in recall of the Hate Speech class and by 3% in macro averaged recall
• Leveraged Hatebase API to expand vocabulary of explicit hate words
• Twitter dataset of 25K+ tri-classed tweets 

Hate Speech and Offensive Language Detection with Bayesian Networks

• Application of Bayesian Networks to the task of fine-grained sentiment analysis to identify relationships amongst features: To detect offensive and hate speech we explored the application of Bayesian Networks to the task of fine-grained sentiment analysis of a set of tweets from Twitter. We found that Bayesian networks perform better than baseline linear classifiers, but worse than a Naive Bayes model in terms of recall for classification of hate speech and conducted an in-depth error analysis to discover why Bayesian Networks performed this way.

PROJECTS 

CS + Social Good: TeachConnect

Team Co-Lead 

TeachConnect is a platform for AP teachers to collaborate and find mentorship (Javascript, ReactJS, CSS, Firebase) 

VMWare Women’s Leadership Innovation Lab Seeds of Change

Cohort Lead

Stanford’s Seeds of Change initiative provides innovative training and support to young women in STEM as they transition through high school and college to successful technology careers.

Led monthly seminar style lectures to a cohort of female HS sophomores working through Stanford’s Clayman Institute Women’s Leadership Modules .

Relevant Coursework 

Programming Methodology; Programming Abstractions; Computer Organization and Systems; Principles of Computer Systems; Probability for Computer Scientists; Natural Language Processing; Natural Language Understanding; AI: Principles and Techniques; Machine Learning; Deep Learning; Computational Logic; Natural Language Processing with Deep Learning; Algorithms Design & Analysis; Software Project Experience with Corporate Sponsors.