There isn’t a single platform where data scientists can seamlessly manage datasets, run Python code, and utilize Driverless AI in one place. As a result, they have to navigate between multiple tabs and tools, which disrupts their workflow and reduces efficiency.
Since we didn’t have clear scope about project and where to start, I thought of using fail fast principle till we build MVP product for this. So we stick with the following process
Discovery → Ideate → prototype → Test
Used the Design Thinking approach for solving this problem. To practice design thinking we followed the below process.
I worked with 3 data scientist in H2O to understand main pain points. We brainstorm the idea and created a document outlining key features and goals of the application. This gave me a brief idea about the application and it’s scope.
To dive into deep of our scope I tried many competitors like Databricks, Deepnote and prepared a document to brainstorm and to collaborate with data scientist to understand the scope.
In Ideate phase I worked on lo-fi wireframes and present and test it out with the data scientists I worked with. I have created couple of concept in this stage to test it out.
Once we had a clear vision and flow about how app needs to work I started working on UI design.
I have created local design system based on components that were unique to this application. All the other components are from out in house design system.
Once we created an MVP version of the prototype I have conducted usability test to identify weather this application is usable for the data scientists.
Some of the feedbacks we received.
| Feedback | Solution |
|---|---|
| Some users didn’t notice the ai prompt bar. It wasn’t obvious enough. | Add user onboarding cards for new users |
| Confused about dataset format. Tried upload a zip file but it didn’t work | display supported file types and prevent uploading other file types |
| Displaying all the actions in all cells | Hide actions when a cell is not selected. |
Although our goal was to create a platform with a single AI prompt, we faced technical challenges in identifying the AI agent. As a result, we had to implement this cell-based AI prompt approach instead.
