Microsoft Research Canvas

Creating a data filtering UI to enable AI use for scientists

My role

3rd designer hired by the team leading team workflows, design systems, core IA, blue sky concepts, and user flows

Deliverables

Detailed production specs for engineers, reusable patterns and frameworks, and sales decks that brought an 8-figure customer acquisition

Team

Product manager
UX design team
Front-end engineers

Timeline

Q3 2020

The what

A feature within Research Canvas that allows users to efficiently perform multiple filter operations on very large tables, unlocking more advanced AI capabilities.

The what

A feature within Research Canvas that allows users to efficiently perform multiple filter operations on very large tables, unlocking more advanced AI capabilities.

The what

A feature within Research Canvas that allows users to efficiently perform multiple filter operations on very large tables, unlocking more advanced AI capabilities.

The why

Users couldn’t efficiently filter their data before running AI models, due to a clunky, limited, and non-scalable filtering experience.

The existing UI lacked basic filter functionality (like range and date filters), required users to write code for common tasks, and didn’t scale to large datasets. This led to frustration, wasted time, and user drop-off, ultimately blocking access to the core value of the product: AI-driven insights.

The why

Users couldn’t efficiently filter their data before running AI models, due to a clunky, limited, and non-scalable filtering experience.

The existing UI lacked basic filter functionality (like range and date filters), required users to write code for common tasks, and didn’t scale to large datasets. This led to frustration, wasted time, and user drop-off, ultimately blocking access to the core value of the product: AI-driven insights.

The why

Users couldn’t efficiently filter their data before running AI models, due to a clunky, limited, and non-scalable filtering experience.

The existing UI lacked basic filter functionality (like range and date filters), required users to write code for common tasks, and didn’t scale to large datasets. This led to frustration, wasted time, and user drop-off, ultimately blocking access to the core value of the product: AI-driven insights.

The how

I led a redesign of our filtering UI. I identified key pain points through user research like missing range/date filters, reliance on code, and poor performance on large datasets. I designed a scalable, no-code filtering interface, validated it through usability testing, and collaborated closely with engineering to ensure performance and flexibility.

The new design significantly reduced filter time and performance, and allowed our product to scale.

The how

I led a redesign of our filtering UI. I identified key pain points through user research like missing range/date filters, reliance on code, and poor performance on large datasets. I designed a scalable, no-code filtering interface, validated it through usability testing, and collaborated closely with engineering to ensure performance and flexibility.

The new design significantly reduced filter time and performance, and allowed our product to scale.

The how

I led a redesign of our filtering UI. I identified key pain points through user research like missing range/date filters, reliance on code, and poor performance on large datasets. I designed a scalable, no-code filtering interface, validated it through usability testing, and collaborated closely with engineering to ensure performance and flexibility.

The new design significantly reduced filter time and performance, and allowed our product to scale.

What research canvas did

Our platform…

helps people understand unstructured data by standardizing, indexing, and unifying siloed data types such as PDFs, images, videos, and CSV though AI.

Novartis, our publicly announced partner, uses our product to aid in the drug discovery and development process.

Our users…

are subject matter experts who lack the technical and data science expertise to develop their own AI models. They are typically researchers, doctors, or analysts.

The old designs

A significant part of our product deals with large tabular datasets such as CSVs. Before users can perform most analysis on tables of 1 million+ rows, they often need to run multiple filters.

During this team’s early days, the priority was to get features shipped. This meant making lots of design compromises. One of the early design decisions the team made was to copy key design patterns from Excel. However, when it came to filtering, once we shipped to our customers, we immediately started receiving feedback that I distilled down into several main pain points :

The old designs

A significant part of our product deals with large tabular datasets such as CSVs. Before users can perform most analysis on tables of 1 million+ rows, they often need to run multiple filters.

During this team’s early days, the priority was to get features shipped. This meant making lots of design compromises. One of the early design decisions the team made was to copy key design patterns from Excel. However, when it came to filtering, once we shipped to our customers, we immediately started receiving feedback that I distilled down into several main pain points :

The old designs

A significant part of our product deals with large tabular datasets such as CSVs. Before users can perform most analysis on tables of 1 million+ rows, they often need to run multiple filters.

During this team’s early days, the priority was to get features shipped. This meant making lots of design compromises. One of the early design decisions the team made was to copy key design patterns from Excel. However, when it came to filtering, once we shipped to our customers, we immediately started receiving feedback that I distilled down into several main pain points :

Users were leaving

Filter UI lacked support for ranges, custom filters for dates, and other expected features, forcing users out of our product

Filter UI lacked support for ranges, custom filters for dates, and other expected features, forcing users out of our product

Time consuming

Both adding and removing filters took users a long time

Both adding and removing filters took users a long time

Required code

Filters types that commonly have a UI in other applications required formulas in ours

Filters types that commonly have a UI in other applications required formulas in ours

Didn't scale

Designs didn’t accommodate large datasets OR additional features

Designs didn’t accommodate large datasets OR additional features

Another observation

The old designs were based on the filtering pattern from Excel, however.

Excel is about letting a user manage their own data. Our product is about consuming and making sense of someone else’s data.

Design exploration

I lead brainstorming sessions, defining the problem space and articulating the core use cases for advanced filtering.

Whiteboarding and on-paper wireframing helped us narrow down on a few promising directions ranging from incremental to large changes.

Design exploration

I lead brainstorming sessions, defining the problem space and articulating the core use cases for advanced filtering.

Whiteboarding and on-paper wireframing helped us narrow down on a few promising directions ranging from incremental to large changes.

Design exploration

I lead brainstorming sessions, defining the problem space and articulating the core use cases for advanced filtering.

Whiteboarding and on-paper wireframing helped us narrow down on a few promising directions ranging from incremental to large changes.

Audit

During my early exploration phase, I discovered a UX pattern called a condition based interface. This pattern is commonly used for filtering UX and enables code-like functionality through a UI.

Audit

During my early exploration phase, I discovered a UX pattern called a condition based interface. This pattern is commonly used for filtering UX and enables code-like functionality through a UI.

Audit

During my early exploration phase, I discovered a UX pattern called a condition based interface. This pattern is commonly used for filtering UX and enables code-like functionality through a UI.

Design direction

Ultimately, the design team, PM, and other stakeholders saw huge potential in the condition-based filtering concept. We decided that the power of this interaction could bring a lot of functionality to our users.

Not only did the designs work for filtering, but also for other table 'modifiers' such as sort, group by, and show and hide columns

Design direction

Ultimately, the design team, PM, and other stakeholders saw huge potential in the condition-based filtering concept. We decided that the power of this interaction could bring a lot of functionality to our users.

Not only did the designs work for filtering, but also for other table 'modifiers' such as sort, group by, and show and hide columns

Design direction

Ultimately, the design team, PM, and other stakeholders saw huge potential in the condition-based filtering concept. We decided that the power of this interaction could bring a lot of functionality to our users.

Not only did the designs work for filtering, but also for other table 'modifiers' such as sort, group by, and show and hide columns

User testing

While the condition-based filtering design had clear advantages, I identified key risks: it could feel too technical, unfamiliar, and require significant engineering effort. Because it was a major shift from existing patterns, I partnered with our user researcher to run a usability study before moving forward.

We tested the new design against a safer, incremental improvement with nine advanced Excel users from pharma and energy - closely aligned with our target audience.

User testing

While the condition-based filtering design had clear advantages, I identified key risks: it could feel too technical, unfamiliar, and require significant engineering effort. Because it was a major shift from existing patterns, I partnered with our user researcher to run a usability study before moving forward.

We tested the new design against a safer, incremental improvement with nine advanced Excel users from pharma and energy - closely aligned with our target audience.

User testing

While the condition-based filtering design had clear advantages, I identified key risks: it could feel too technical, unfamiliar, and require significant engineering effort. Because it was a major shift from existing patterns, I partnered with our user researcher to run a usability study before moving forward.

We tested the new design against a safer, incremental improvement with nine advanced Excel users from pharma and energy - closely aligned with our target audience.

User testing results

The new Advanced Filtering designs were very well received and preferred over the current filtering design by all participants. Participants found the new designs to be more efficient and visually appealing than the current design.

“This makes much more sense in an everyday use case.”
“I don't have to move around and click. It's all in one place”
“You have so much more control of your filters”

The final designs

Based on strongly positive user feedback, we felt confident moving forward with the design. The research findings gave the design team the evidence we needed to make the commitment to introducing the new filtering pattern. I have outlined a few key features below:

The final designs

Based on strongly positive user feedback, we felt confident moving forward with the design. The research findings gave the design team the evidence we needed to make the commitment to introducing the new filtering pattern. I have outlined a few key features below:

The final designs

Based on strongly positive user feedback, we felt confident moving forward with the design. The research findings gave the design team the evidence we needed to make the commitment to introducing the new filtering pattern. I have outlined a few key features below:

Scaling for additional features

I didn’t just create designs for advanced filtering. I applied the same patterns consistently to other table modifiers such as group by and sort by.

Expansion to more advanced AI

Since our product is AI-focused, I created blue-sky concepts for how table filters might be expanded to include features such as image recognition, format conversions, and entity classification.

Impact

Although at the time I left, our product had not been released to the general public public, we have customer usage at Novartis as well as our other partners. Overall, my biggest areas of impact with this project have been:

  • Very positive customer feedback - enabling users to complete tasks and workflows that were previously impossible with the old designs

  • Defining a design pattern that has been used for other interactions by other members of the design team

  • Setting our product up for success as we scale up to accommodate larger data sets and more diverse use cases

©️ Owen Whiting, 2025 all rights reserved

Contact

LinkedIn

Email

Instagram

©️ Owen Whiting, 2025 all rights reserved

Contact

LinkedIn

Email

Instagram