Citadel

• Date: Summer 2019
• Technologies: C++, Python

During the summer of 2019 I worked at Citadel Asset Management on the Global Quantitative Strategies Core Technology Group.

Autolab

• Date: Spring 2019
• Technologies: AI/ML, TensorFlow

Coming Soon.

Modelzoo.live

• Date: Spring 2019
• Technologies: Clipper, AI/ML, Docker, TensorFlow, Keras

I spearheaded the model curation team for Berkeley RISE lab's Modelzoo.live project. The project is designed to highlight the ease of hosting machine learning models with the Clipper framework. I modified, designed, or trained several deep learning models to be hosted on the Clipper Platform including ResNet, InceptionV3, text sentiment RNNs, and Tensorflow Object Detection MobileNetV2 etc. and created custom docker containers to run them. The website is currently in beta testing but was demoed at the 2019 Spark - AI Summit. You can find it here.

Launchpad - Dream RLs

• Date: Spring 2019
• Technologies: AI/ML, TensorFlow, Keras, Reinforcement Learning

For my launchpad project in Spring 2019, I worked on a team implementing and experimenting with the Worlds Models Machine Learning paper found here As a team we implemented various aspects of the paper. I was responsible for working on the MDN-RNN, or mixed density network recurrent neural network. It's purpose was to predict future states of the environment (in the form of the latent vector output by the variational autoencoder) given the previous state. Above is an outline of how the overall architecture works, and below are two examples of our work. First is a reconstruction of the Open AI Gym Lunar landing game environment. The second is an example of an MDN-RNN predicting future states of the Open AI Space Invaders game.

Artifical Intelligence Internship at Intel

• Date: Summer 2018
• Technologies: AI/ML, TensorFlow, Keras, AWS, Computer Vision

Over the 2018 summer I worked as an intern at the office of the CTO at Intel's Artificial Intelligence Products Group (AIPG). During the span of my internship, I worked on a variety of different projects. During the first portion of my internship, I created a demo using Intel's OpenVino Model Optimizer and the AWS DeepLens, powered by Intel's X5 Atom processor. I trained an image classification network to sort photos of waste into over twenty categories by type, then from there into recycle, compost, or garbage. After that, I built a natively running application on the DeepLens with a GradCam gradient visualization to identify the ResNet50's regions of high attention. The app let users provide input to collect more data. I wrote a tutorial for building such systems, which can be found here.

While creating the demo system, I worked on documenting the AWS machine learning workflow for Intel. I trained models with EC2 instances and S3 backed datasets before moving to explore SageMaker. With AWS SageMaker, I designed workflows for training custom TensorFlow and Keras models with custom scripts. I used job dispatching to also tune model hyperparameters.

For the final portion of my internship, I explored gradient based model explanation systems, particuarly for NLP and image classification tasks. I'm currently still working on this, and will provide updates when done!

Launchpad - Unique Object Tracking

• Date: Spring 2018
• Technologies: Python, Scikit, YOLOnet
• Role: Primary Project Developer

Launchpad is a UC Berkeley Student run club that pursues various research style projects in AI and machine learning. During the Spring 2018 Semester, I worked on unique object tracking through video. Rather than just identify objects, our object was to track their uniqueness through time. I developed algorithms to apply nearest neighboring clustering to frame buffers to group objects together, then later applied regression and ML models to predict their future locations. We experimented with multiple regression based models and with a function predicting nueral network trained in Keras. Our final product can be seen in the video above. Object IDs are kept consistent through frames.

Switchboard App

• Date: Spring/Summer 2018
• Technologies: Java, Android, SocketIO, Rest APIs, Tokbox API
• Role: Android Developer, Responsible for backend connections

As part of the contract team for the Mobile Developers of Berkeley during the Spring 2018 Semester, I developed the Android variant of Switchboard, a mobile communication channels application owned by the Switchboard SkyDeck startup. The premise of the app is to have a persisent walkie-talkie like connection to all of your contacts or co-workers. To chat with others, you can simply tap their name to start streaming your audio to your shared channel. While working in a team of three, I handled most of the backend interaction. I interfaced the actual voicechat audio streams with the TokBox restful API. I managed connections to a socketIO server for pulling updates. Finally, I implemented calls to the Switchboard team's custom API.

Atmospheric Measurement System Development

• Date: Spring 2018
• Technologies: Python, SQL, Raspberry Pi, I2C

I can't release my report for this project online, as the research work was for the ROMPS group at UC Berkeley. As a researcher at the ROMPS group, I was working on a project proposal for an oscillating atmospheric probe to study low hanging cumulus clouds at about 100m of elevation. The basic premise was to create a custom sensor package whose height could be measured and controlled over a long period of time to collect previously unobtainable data. I implemented the data collection procedures using a raspberry pi with I2C temperature, pressure, and humidity sensors that pushed data to a mySQL database. I designed the basics of the winch system and balloon like devices after extensive research and physics calculations.

Streamline App

• Date: Fall 2017
• Technologies: Android Studio, Java, Firebase, Spotify SDK
• Role: Android Project Manager

Streamline is a cross-platform mobile application developed by me and other members of the Mobile Developers of Berkeley Organization. The application allows users to share music by searching the entire spotify library and then posting one song to a feed each day. As project manager, I lead android development for our team of three coders. During the course of the project I worked with Token authentication, Firebase, and Android Studio. The app is availible on the playstore at the above link. As a part of the team I designed the app screens, implemented authenticaiton and streaming features with the Spotify API, and read / wrote to the Firebase database.

Language Processing for SEC Filings

• Date: Summer 2016
• Technologies: Python Natural Language Toolkit, SQL, Javascript, Google Charts API
• Role: Research Assistant at Stanford Department of Political Science

Over the summer I worked as research assistant at Stanford's Department of Political science under the supervision of assistant professor Andrew Hall. I worked on a research problem involving 1.3 Terabytes of 10-K and 10-Q filings that were scraped from the U.S. Securities and Exchange Commission. The basic approach was to format, parse, and then split the text files into n-grams, units of varying numbers of words, in order to make statistical correlations against time, industry, or firm. I completed all natural language processing in the natural language toolkit python environment. The end goal of the project is to create a publicly available web visualizer for all of the language data, similar to the Google N-Grams visualizer for all texts within Google Books.I was as the primary programmer on the research team of approximately eight graduate and undergraduate students. I developed the algorithms to organize and interpret the raw text of the scraped corporate filings. The first step was to use python scripts in order to preprocesses and clean the files, before accurately parsing information from them. From their I completed two different objectives. I designed a robust program flow designed to extract common bigrams from all of a single firm's filings, essentially resulting in a sparse frequency matrix of all firms. My second task was creating the web interface for performing searches and visualizing results for the word tokens in the filings. Users can plot word frequencies over time and against industry segments. This involves a full-stack with an multi-table indexed SQL backend in conjunction with PHP, HTML/CSS for the site, and JS and Google Charts API for visualization. I wrote all of the code, defined the SQL schema, and spent hours optimizing the runtime.

FRC Robotics 2017

• Date: Spring 2017
• Technologies: Solidworks, Mill & Lathe Fabrication, C++
• Role: Team Captain, Manager

I served as captain during the 2017 season for my high school's robotics team. Under my direciton, our team earned first place at the Venture Regional in California and qualified and competed at the FRC naitonal championship for the first time in our school's two-decade history. As captain I managed the build process for the robot from ideation to designed and refinement. Using 3D solidworks I designed over 200 components, completely designing the robot's drive and game piece control systems. I manufactured parts in-house utilizing a milling machine, Lathe, and 3D printer and additionally interfaced with outside professional manufacturers. The robot moves in any direction at 20 ft/s, collects and places gears on pegs, and is able to hang itself from a rope with a knot.

One of the robots most complicated systems is its "swerve" drivetrain. Using a complex system of gears, motors, and continuous potentiometers, the robot is able to rotate each one of its wheels independently while keeping them under power, allowing it to move in any direction without rotating the robot itself. Renders of the CAD I made for it are pictured above.

FRC Robotics 2016

• Date: Spring 2016
• Technologies: Solidworks, Mill & Lathe Fabrication, C++
• Role: Team President, Manager

During the 2016 season, I worked on a complicated robot designed to traverse obstacles and shoot a foam ball into a goal about 11ft in the air. Through rigourous prototyping, I designed a drivetrain and frame with pnuematic wheels that could not get stuck on the game ostacles. In addition, I designed a pnuematically actuated flipper system in order to allow the robot to traverse the obstacles that required active action by the robot. Finally, I worked on implementing the aiming system for the fly-wheel shooter. As team President I managed a team of over thrity students throughout the build process.

Language Processing For Speech Sentiment

• Date: Spring 2017
• Technologies: Python, Natural Language Toolkit

For my senior project I decided to research the changing negativity of political speech overtime using natural language processing. Utilizing the NLTK python library, I wrote code to measure the negativity of text phrases using the VADER speech classification algorithm. From there, I split speeches up into their individual sentences and then was able to tokenize each sentence and assign it a negativity score before averaging all sentences to reach a file score. Eventually, I was able to plot the relationship between time and the negativity of speech using the presidential Innaugural address and State of the Union address. I found statistically significant results that indidcated a negative trend of speech sentiment.

Later, I analyzed the political debates specifically. I wrote a python script to parse through the text of the debates between Clinton and Trump and divide the text by speaker. I later compared the sentiment scores of each speaker to determine if one candidate was more negative than the other. My results indicate that Trump was 312% more negative than Clinton.

Web Development

• Date: 2015-Present
• Technologies: HTML, CSS, JavaScript, jQuery, PHP, Bootstrap
• Role: Android Project Manager

In 2015 I created a new website for the non-profit Beyond Fistula (now a legacy site), a charity that assists women in Africa after invasive fistula removal surgery. As web manager I updated webpages, styles, and wrote content for the site which is used in the newer version. I later took CS193C at Stanford University, Clientside Internet technologies, where I learned JavaScript. In the class I completed various projects, one of which was to make a rudimentary Google Maps like clone in JavaScript. Afterwards, I began to use web development skills in several projects and eventually created the portfolio website you are on now.