- CLIENT: Microsoft AI Architecture & Strategy
- ROLE: Product Design
- Conceptual Design
- User Research
- Personas & User Flows
- Information Architecture
- Heuristic Assessment
- Wireframing
- Rapid Prototyping
- User Testing
- Style Guide & Hi-Fi Comps
- Rapid Prototyping
- Front End Development
- Delivery and Integration Support
Overview
One of the primary deliverables for the AI Architecture & Strategy team at Microsoft was to rebuild the Model Training tool. Within the context of AI, one of the interesting things is that humans train the AI to complete the right behavior within a set of parameters. The Model Training tool, as it was aptly named, was the process in which an AI model was created for use.
Problems & Opportunities
However, the tool itself was difficult to understand and use, resulting in losing valuable time waiting for these AI models to be ready for use in another product. I was selected to lead design efforts to redesign this tool so that it's easy to use, will complement other tools within AI Architecture & Strategy team and most importantly, speed up AI model creation to scale for immediate and future needs.
Roles & Responsibilities
As the lead designer on this project, I worked with a fantastic dev team, a great PM, and sought the input of the design & dev team for feedback, especially to a/b test, to deliver this tool. The main goals were to bring the the tool up to MSFT design standards via Fluent Design, make the tool easy to use for both the primary users of the Model Training tool, the people training the AI models, & secondary users, the Data Scientists. Give all users a holistic view of the tasks required, and status. Finally, redesign the tool in such a way that it can scale modularly as more AI models are added.
Selection of AI Models
Below are a few examples of some of the models being trained - Annotating .PDF documents where the model identifies parts of a document, Polygon Annotation - whether identifying multiple people or objects in photos or video, Relational Extraction which is pulling out of a document related relationships between the content in the document, Sentence Labelling which is an assessment of a given sentence and how it relates back to the context of the document, and lastly Table Q&A, where the model creates a questionnaire based the context of a table of content.
This is just the beginning phases of some of the models that will be trained using this tool. As the final product will be open ended, it was important to think about how this tool would be both modular and scale over time.
Starting Point
The initial tool was strictly functional, both in terms of design and usage. It was imperative to have AI models built out ASAP as a dependency for another program that was being built. A talented front end dev was able to build this out very quickly.
Research
I spent a good deal of time on researching what other competitors were doing both manually, and with AI in terms of entity & sentiment linking, data annotation and a few other common AI tasks. It was fascinating to see the different approaches taken and these discoveries were a stepping stone to my own ideas about how to best accomplish this task for the customer.
Design Breakdown
Breaking down the problem into solving for the needs of the data taggers, the data scientists & managers here revealed a third need: an admin to be able to have control over the system including all the capabilities of the data scientists. Breaking each persona out by task was imperative to understand and from there, I was able to build out a flow (above) for all parties.
Component Design & Function Mapping
As the backend was complete and remained 80% unchanged the process, this neccessitated function mapping where I reorganized the front end to be more in line with both the Fluent UI design language and more accessible and easier to use for the customer. This allowed dev to easily see where the new functions were and how best to realign.
I also spent a lot of time finding the right color scheme, developing custom components and testing out different layouts, then a/b testing in small groups with the greater dev & design team. This resulted in aligning with the initial pitch of a user dashboard for all users to see, giving the user a high level overview of the current workload at any given time.
End Result
The end result is a beautiful product that's easy to find any information, easy to use, aligns to Fluent UI, & is more intuitive for model trainers, data scientists & and admins. With the new dashboard, it's easy to see what is on the plate for both the model trainers, & data scientists, giving all users a holistic view of the tasks required, and status. With the current backbone in place, adding new AI model tasks is simple and organized, allowing this tool to scale as needed for AI Architecture & Strategy at Microsoft.
After this was launched, several people on the team reached out in congratulations and thanks for transforming the tool into something more useful, saving them time & effort towards more pressing pursuits. My favorite part of designing for internal tools is probably that last bit - getting instant feedback from the 'customers', - my coworkers.
Chad Rawlinson
Product Designer
Microsoft AI Research & Strategy
"Gaining a deeper knowledge of AI and how it impacts building better tools has been inpsiring. I am proud to give my customers, the data scientists better tools to work with. Having seen how AI models behave on the back end and how they are developed fills me with optimism towards a future where research results are exponentially increased. It's been a fantastic project to work on with AI Architecture & Strategy, and this tool has laid the groundwork towards how to utilize these AI models to provide intelligent systems that deliver incredible results."