The SUMSarizer is the brainchild of three Berkeley researchers who found their work studying improved cookstove usage in developing countries severly hampered by tedious data analysis. They solved this problem with a clever application of machine learning, and I helped them make this tool available to the wider research community.
I helped them get over the initial hurdles of building a user-ready web application, aligning their goals realistically, and setting up the architecture for their small team of graduate students to add software features.
An initial week-long engagement provided them enough steam to work for nearly a year, ultimately launching the tool to a group of other researchers in the field.
The tool received fantastic reviews, and we are currently working together to get it ready to launch to an even wider audience.
The Open Source collaboration project was thrilling to see succeed: a bit of expert effort and the right structure allowed a tool that would have otherwise languished in the lab to impact teams around the world, ultimately helping better serve communities in need worldwide.
User Interface and Backend for Machine Learning Training
The SUMSarizer app allows researchers to label events measured by temperature sensors (called SUMS) attached to cooking stoves. I integrated the prototype D3 visualization with a Flask backend and built the core CRUD features.
I thoroughly documented the web app setup to provide the framework for two graduate students to flesh out the rest of the app. They worked gradually over about a year without any assistance.
The app was used to conduct an initial study of human vs. machine learning performance using a group of fourteen researchers from labs and NGOs. The results were quite encouraging, showing high accuracy and low bias compared to by-hand labelling.