This is a high-level summary of my work due to NDA. Please feel free to contact me at cll246@cornell.edu to learn more. :)
Simplifying Machine Learning pipelines for data scientists
Internship
COMPANY
↳ The Home Depot
ROLE
Sole UX Designer
TEAMMATES
2 SWE Interns
6 engineers
2 stakeholders
TIMELINE
May - August 2022
OVERVIEW
Blueprints
On the Machine Learning (ML) Enablement Team, I worked cross-functionally to improve Blueprints, a machine learning (ML) drag-and-drop workflow that helps data scientists visualize and compute data. Through UXR, design studios, concept tests, usability tests, and sprints, I was able to increase Blueprints's task success rate from 0 to 100% and decrease the time to complete tasks by 92%. Blueprints shipped by the end of my internship and is still being iterated on today.
Problem
Not a single data scientist is able to build and edit a ML pipeline.
Business Goal
Data scientists could save a ton of time and money by using reusable components to build pipelines.
Solution
Redesign Blueprints to ensure data scientists could create and visualize pipelines easily with a myriad of new features (under NDA).
HOW ML PIPELINES ARE MADE
Components are the building blocks of a ML pipeline.
This is an example ML pipeline. Pipelines can be made of just one component, or even tens!
This is part of a component. Even just one can contain a lot of code!
PROBLEM
Data scientists' (DS) woes
Increased cognitive load
When components are listed as code, it's hard for data scientists to visualize the overall pipeline.
Slows down workflow
Without a connected workflow, DS must switch between local code environments to running models on the cloud, which is tedious and time consuming.
Adds friction among DS
Since components are hard coded, every data scientist created their own components with their own code, leading to a lack of standardization.
SAVING MONEY
It can take months to make a pipeline.
Faster pipeline building means more money saved.
THE SOLUTION
Blueprints:
[ˈbluːprɪnts] noun
a machine learning drag-and-drop workflow that helps data scientists visualize and compute data.
USABILITY ISSUES
User interviews and usability tests show the original Blueprints is highly unusable.
I conducted 6 user interviews and usability tests with DS of differing teams and experience levels. Tasks were barely completed, even with guidance.
MY DESIGN & RESEARCH PROCESS
Many stakeholder meetings, prioritizations, affinity maps, design studios, concept & usability testing, and iterations later...
With as much detail under the NDA, I worked with stakeholders to build prioritization matrices and affinity maps to help prioritize which aspects of Blueprints to focus on. We ended up choosing pre-built component customization as the focus of the project, due to its high user impact and relative ease of implementation. Through many rounds of iteration, design studios, and testing, I had created final designs ready to be shipped.
RESULTS
100% success rate in task completion, 92% decrease in task completion time
What used to take data scientists minutes on end, now only took a few seconds. DS said the new Blueprints would standardize the way they work with others, help them abstract and visualize code, and be a training tool and resource for others.
NEXT STEPS
Blueprints is in production!
We're proud to announce that Blueprints is in production! The engineering interns and software engineers did a great job of implementing the design changes so that we could release the new features in our MVP to immediate "friends and family". In my last week of my internship, I onboarded an incoming full-time designer who would take over my work on Blueprints.
LESSONS
I learned design & soft skills
I am very proud of the work and growth I accomplished in my first UX internship, and I'm happy to continue my work with THD in the Fall with their OrangeWorks Innovation Labs program, where I will be doing UX in the Virtual Reality space!
Adaptability
In the beginning, we focused on the customization of pre-built components. However, I saw other low-hanging fruit that wouldn’t affect the timeline and were a value add, so I proactively took on these designs.
Working with constraints
We were on a tight deadline and engineers wanted to start implementing before some of my designs were tested. I let engineers code aspects of the design that I knew were not going to change, while I conducted my tests and finalized designs.
Storytelling
I was fortunate enough to practice my report-outs every week, and eventually presented to an audience of UX managers, stakeholders, ML engineers, and even the VP of The Home Depot! Getting feedback from different audiences helped me craft my storytelling.
This is a high-level summary of my work due to NDA. Please feel free to contact me at cll246@cornell.edu to learn more. :)