About Patrick Russo


I'm a data scientist/machine learning engineer based in New York and Co-Founder of OpenAxis.
I've worked at a few startups, including as a machine learning engineer at theMednet and a data scientist at Via Transportation. Before that, I was a Senior Research Analyst at the Federal Reserve Bank of New York (in the International Research Function) where I worked on macro-finance research. I graduated from Connecticut College in 2014, majoring in economics and mathematics (with a concentration in statistics).

My contact information is on the left, my projects and publications are below.



To avoid dealing with more HTML than needed, I'm going to start being lazy and posting some short posts on Medium. So far, the posts are just some random AWS notes or brief guides to do obscure AWS stuff - enjoy! https://medium.com/@patrusso2

Notes include:

Triggering a Lambda using SNS from a different Region in the Serverless Framework

Protect your single region API gateway using AWS WAF

Point Custom Domain to an AWS Gateway




Recent Research


OpenAxis

Easily visualize, share, and collaborate with data!
See OpenAxis.


80-20 News

A platform to bring you 80% of the news in 20% of the time! Scraping 40k articles a day and producing abstractive summaries using PEGASUS.
See 80-20 News.
Or see the live demo (please bear with me, the prototype takes 10 seconds to load). Feed updated each morning.


Online Cookbook

It's not actually BERT, but Cooking With GPT-2 sounded a little more awkward... React and next.js on lambda
See Cooking With Bert here!




iOS app with python/lambda backend

I've been working on an iOS app with a python backend running on AWS Lambda. It uses serverless Aurora DB so it's entirely serverless and scalable.

I'm not talking much about the app - but have been posting some things that I've learned on Medium (see above!).





iOS Object Recognition

I have little experience with convolutional neural networks and I had never previously used Swift (or any other App programming language). So, I decided to jump in head first and try them both together. I put together a small app that reads photos from your library (or lets you take new photos) and then posts that photo to a Flask webserver. The webserver then runs the photo through a trained CNN and produces a label and returns that label. The CNN I am training is using ILSVRC2012 and is based upon GoogLeNet (former versions were based on AlexNet, ZF-Net, and Vgg).


I ended up with about 65% top-1 prediction accuracy with GoogLeNet. (I'm lazy and left out the auxiliary classifiers as for my purposes an improvement of 0.5% seemed negligible.) I haven't measured the top-5 accuracy yet, but it definitely feels pretty accurate - and mistakes are quite reasonable (e.g, a clear glass vs. a measuring cup or the back of a frying pan hanging against a wall vs. a gong.)

Some tips are in Tensorflow on the left, including a few pitfalls I ran into setting up and training a CNN from scratch.

If you want to get into CNN's I'd recommend Andrew Ng's new Coursera courses and/or reading the articles here (the articles themselves provide much more detail than the summaries): https://adeshpande3.github.io/adeshpande3.github.io/The-9-Deep-Learning-Papers-You-Need-To-Know-About.html

A few example images are here (the app itself is clunky but was built in 2-3 days as I learned Swift):






Twitter: Clustering with K-Means

I use Python to download tweets from 60+ news organizations, pundits, and politicians and cluster them automatically into ≈13 groups based upon the words in the tweet.  This is updated throughout the day (unless there is trouble with the internet connection).  More information can be found on the twitter page).

Description Results



Primary Debates: Live Transcripts

I have recently been working on building a speaker classifier to detect the identification of whoever is speaking.  This prediction is joined with the content prediction from the Google Speech API (so that I can predict what was said and who said it) to create a live transcript during debates.  This is still a work in progress, at this point I'm just posting it for fun, but it still needs more work.  

See the transcripts that were produced during the last several debates



New York Times Articles: Recommendation System

Using a tf-idf bag-of-words approach I use the pairwise comparisons between the text of all articles on a given day to determine the ranking of which articles would me most interesting to me.  

See the most recent article recommendations




Publications

Federal Reserve Bank of New York

Blog Posts

Lower Oil Prices and U.S. Economic Activity. With Jan Groen, Liberty Street Economics, May 2016.

Is Cheaper Oil Good News or Bad News for U.S. Economy? With Jan Groen, Liberty Street Economics, June 2015.

The Myth of First-Quarter Residual Seasonality. With Jan Groen, Liberty Street Economics, June 2015.

Falling Oil Prices and Global Saving. With Tom Klitgaard, Liberty Street Economics, June 2015.


Other Reports

Oil Price Dynamics Report -- a weekly oil price decomposition update (based on the blogs with Jan Groen) 

Global Economic Indicators  -- Developed an R script to produce PDF booklets of charts (20+ booklets, 600+ charts) on a daily basis (the link is just a public example)



American Institutes for Research

Gandhi, Allison Gruner, Rachel Slama, So Jung Park, Patrick Russo, Kendra Winner, Robin Bzura, Wehmah Jones, and Sandra Williamson. "Focusing on the Whole Student: An Evaluation of Massachusetts' Wraparound Zone Initiative." Journal of Research on Educational Effectiveness just-accepted (2017).


Connecticut College

Honors Thesis
Determinants of Undergraduate GPA and Persistence at Connecticut College, May 2014.
Supervised by Terry-Ann Craigie, Ph.D., Candace Howes, Ph.D., and John Nugent, Ph.D.

Teaching & Learning Newsletter, page 18. Connecticut College:
Peers, Academic Performance & Persistence Beyond the First Year, Spring, 2015.



Further Information & Contact Info:


View Patrick Russo's profile on LinkedIn

E-mail: patrusso2@gmail.com