How to make tracking ML data as simple as tracking a UPS delivery?
Connecting
Disjoint
Workflows as One
ROLE
Product Designer
DATE
Match 2020
- June 2020
PROJECT STATUS
🚀 Launched
FEATURE TEAM
Ricky Zhang (Manager), Maggie Li(PM Intern)
+ team of 6 software engineers
Project Context
TuSimple develops the world’s most advanced self-driving technologies specifically designed to meet the unique demands of heavy-duty trucks. For self-driving companies like TuSimple, ML(Machine Learning) takes up a heavy part of their workload. In a typical ML workflow, the collection of raw data is one of the core scenarios. Wayzard is the product based on the demand improving the raw data collecting workflow.
In TuSimple, we have about
-
50 + truck drivers
-
50 + Testing Operators (TestOps)
-
100 + data engineers
-
200 + data labelers
They work on the testing routes in California, Texas, New Mexico, China, and Sweden, collecting the fresh blood for Machine Learning. They are the primary users of Wayzard.
Let's See How They Work Together
Workflow Overview
What's the Problem?
User Insights
from Feedback to User Needs
We analyzed users’ feedback and summarized it into crucial User Needs.
-
There should be a place to see the entire data lifecycle.
-
Provide TestOps and drives more handy HCI processes in their working scenes.
-
Access to quickly tracking and responding when an error happens
Analysis
We find that the current workflow is disjointed, which caused most of the current pain points.
Tasks are scattered across five different services at different stages, making work handover and tracking of work a big problem.
How to make tracking ML data
as simple as
tracking a UPS delivery?
Solution
Ideation
Create a unified platform that can
-
Display the entire lifecycle of a data package. From Demands to Delivery
-
Provide a One-Stop style workflow for TestOps and Drives.
-
Quickly pin any roles, for any error, at any stage.
.png)
Design Process
Develop a product from the very beginning, exploring all possibilities.
First Step
Highlight Key Scenarios
I don’t want to significantly change the current workflow, letting my colleagues feel no-struggle to learn how to do their job on the new platform.
Second Step
Designing Information Architecture
Each role has its work zone. They do their jobs together but don’t interrupt each other.
Third Step
Prototype Developing
ML PMs don’t need to hand out tasks on DingTalk and update their requirements through Slack or any platform. It’s more reliable to do this by fill in a form, which requires them to fill their all requirements at once.
Before
After
ML PMs don’t need to hand out tasks on DingTalk and update their requirements through Slack or any platform. It’s more reliable to do this by fill in a form, which requires them to fill their all requirements at once.
Before
After
I designed pretty simple interactions for TestOps. What they need to do is just tapping several buttons. Suppose things are going wrong, just one tap to Slack on-call SREs or the person in charge automatically.
Before
After
Impacts
After the launch of the first public version of Wayzard, we received much appreciation. So Infra decided to discontinue maintaining and updating the MMS, VTS, and VM services one after another, and users also stopped using DingTalk and Slack for task handover.
The entire data collection workflow was shifted to Wayzard, and the efficiency and success rate of data transfer increased by more than 50%.
For more information, feel free to check out the feature post and documentations.
.png)
✌️✌️
Next Steps
Let automation do more.
Some operations are unnecessary if we can add more automatic processes into this Wayzard. It would improve the efficiency and user experience of my product.
Support for multiple devices
They were planning to let the Vehicle-carried computers and Tablet replace some of the laptop’s work. It’s easier to operate especially in a truck or any of the environments out of the office. It would be different. We can use many more convenient interactions such as swiping, double-clicking, etc.
More detailed order classification feature
It can help Testing PMs identify which orders can be put into one trip and increase the tasks one trip can do. It is an excellent way to improve the testing trip efficiency, helping us to collect more raw data and finally help trucks self-drive better.