Joshua Shapiro

About Me

I am currently a Staff Machine Learning Engineer at AKASA focusing on how to leverage natural language processing to solve the systemic inneficiencies in America's healthcare systems. I lead the machine learning development of our medical coding product, focusing my attention on decreasing the time it takes to productionize research results. We're always looking for new people to join our team -- check out our website here.

I also teach the Computer Science Capstone course at The George Washington University, where I guide students through software development projects while emphasizing industry best practices.

Previously I worked at ASAPP both as a Research Engineer and Machine Learning Engineer focusing on natural language processing problems related to customer service.

Before ASAPP I worked at IBM Research in Yorktown Heights, NY. There I discovered my passion for deep learning while working on video action recognition. I worked in the data centric systems department focusing on the intersection of machine learning and high performance computing. I had the privelge on working on two very different projects: one involving computer vision and the other involving temporal clustering of lipids in molecular dynamics simulations.

I graduated in May 2017 from The George Washington University with a BS in Computer Science, and spent a semester of my junior year studying abroad at Korea University.

When I'm not writing code or thinking about machine learning, I enjoy rock climbing and playing piano.

Education

I completed my Bachelors of Science in Computer Science at The George Washington University, graduating suma cum laude. In my undergrad I participated in the honors program and held executive positions in the GW chapter of the ACM.
In my sophomore year, I developed an interest in entrepreneurship, and had the opportunity to work for the university's New Venture Competition. I also interned for Pedal Forward, a B Corp focused on building sustainable bicycles out of bamboo. I went to the Rice Business Plan Competition on behalf of Pedal Forward and won $10,000 for best sustainable venture.
In my junior and senior years I was an undergraduate teaching fellow for Computer Architecture, Software Engineering, and Algorithms 2. I also spent these two years interning part-time with IBM working in their Cloud division.

As a Clark Scholar I was required to study abroad for a semester. I chose to study at Korea University in Seoul, South Korea. It was one of my favorite parts of college as I was able to explore so many new places I never dreamed I would have the chance to visit.
As an elective I took a korean linguistics class which focused on how sounds are formed for the korean language. This was the first time I was exposed to looking at language in a scientific way, and it opened my eyes to the field of computational lingustics and natural language processing.
I hope to return to Korea sometime in the future, as it feels like a second home.

In a less formal setting, I enrolled in a handful of coursera courses between October 2017 and May 2018. I completed these in my personal time while working in IBM Research. Please find the list of courses below. Together the last 5 encompass the Coursera Deep Learning Specialization.

• Machine Learning | October 2017
• Neural Networks and Deep Learning | January 2018
• Structuring Machine Learning Projects | February 2018
• Improving Deep Neural Networks: Hyperparameter tuning, Regularization, and Optimization | February 2018
• Convolutional Neural Networks | April 2018
• Sequence Models | May 2018

Employment

Resume

AKASA is an AI startup focused on solving the systemic inneficiencies the arise between hospitals and insurance companies in America's healthcare systems. I focus on how we can best leverage nlp technologies to achieve our goals. Currently I lead the machine learning development for AKASA's medical coding solution. During my time as a Senior Machine Learning Engineer and Staff Machine Learning Engineer, I've worked on all pieces of our ML pipeline, including developing the tooling and infrastructure to support distributed model training, running LLM experiments and analyzing results, building out standardized inference artifacts, and determining how we deploy > 10B parameter models in production.

I teach the Senior Design capstone course for the Computer Science department. Across two semesters, I guide students through end-to-end software development projects, while emphasizing industry best-practices.

I worked for an AI startup focused on creating innovative AI-native solutions for enterprise. Right now we're focused in the NLP space, revolutionizing customer service interactions.
I started as a Machine Learning Engineer where I brought research results into production. I implemented new services focused on classification and conversation summarization, and later designed and implemented an entity recognition service for dialogue systems. This relied on custom NER models, 3rd party libraries like Duckling, and heuristic approaches to identify, extract, and normalize both generic and domain-specific entities.
As a Research Engineer I primarily focused on our language modeling initiative to generate rich conversational embeddings, decreasing the need for annotated data and increasing performance across a variety of production models. I've researched using a novel attention-based RNN architecture for hybrid ASR and applying the results of that work into production. You can read more about it here!

I worked in the data centric systems division at the intersection of high performance computing and deep learning. During my time in IBM Watson Research I focused on two dowmains: computer vision and molecular dymanics.
In computer vision I worked with a team developing novel techniques for highly scalable video action classification. The goal of the project was not only to get highly accurate results, but also to train on massive amounts of data quickly. We used convolutional neural networks and network-computed optical flows to train an action recognition classifier in parallel on 16 gpus. We developed a new optimization technique for parallel training called AAVG, which allowed us to achieve state of the art performance on UCF-101 when using 2D CNNs. For more information please look at the publication here. If you have any questions about this work please feel free to reach out!
I also worked on a molecular dynamics project collaborating with Oak Ridge National Labs. As part of a larger drug discovery application, I worked on developing novel temporal clustering techniques for lipids in molecular dynamics simulations. I prototyped a density-based clustering system that could find and visualize lipid rafts in a lipid bilayer. This work is still being researched and developed.

I started interning for IBM Cloud Managed Services (CMS) during the summer of my sophomore year. I was then able to continue my internship throughout my junior and senior year working part time remotely, and I returned to Minnesota my junior summer to intern fulltime. Over the two years that I interned in the cloud division I developed internal toolsets to aid in automation and continuous integration. This included creating report generation tools for the CMS project lifecycle management platform (Rational Team Concert), which enabled a more accurate measure of progress and a greater efficiency in allocating resources. This was my first introduction to agile programming methodologies, and I learned a lot about project management working with individuals all across the globe.

I worked as an undergraduate teaching fellow (George Washington University's title for undergraduate TA) during my junior and senior year. I assisted in classes and labs for Computer Architecture, Software Engineering, and Algorithms II. My duties included leading weekly study halls and tutoring sessions, assist the professors in leading students through course exercises, and hosting exam review sessions to facilitate student success.

During college I explored my interests in entrepreneurship by interning with a socially concious startup called Pedal Forward. The ogranization focused on building bicyles out of sustainable bamboo, and employed the homeless in New York to manufacture the bicyles.
I assisted in developing Pedal Forward's Kickstarter campaign to kick off their marketing strategy. Additionally I competed on behalf of the company in the 2015 Rice Business Plan Competition, winning $10,000 for Best Social Venture.

In addition to working for Pedal Forward, I worked for my university's New Venture Competition. There I developed an archive of previous years' competition finalists to better understand the full impact of the competition, as well as assisted in data collection of over 100 teams that entered during the 2015 year.

In my sophomore year assisted in developing and programming an inexpensive quadruped robot power by a Raspberry Pi. This involved using Python to program Dynamixel servos to control the quadruped's movement, and the end gaol was to use external sensors to have the quadruped autonomously walk through its environment.
This was my first experience in a research lab, and though my project was not a success it was instrumental in my decision to study machine learning.

I designed a mockup user interface for an OS X program focusing on personal organization and password storage. This included developing a prototype for a random password generator that was eventually implemented into the program.

During my summer before my first year of college I interned at a tech company focused on auction management solutions. I fixed bugs and updated data for the customer facing websites using PHP, SQL, and JavaScript.

Skills

Deep Learning

Programming Languages & Tools

Contact

• Twitter
• Facebook
• Instagram
• GitHub

Elements

i = 0;

while (!deck.isInOrder()) {
    print 'Iteration ' + i;
    deck.shuffle();
    i++;
}

print 'It took ' + i + ' iterations to sort the deck.';