AI Workbench

Overview

 

AI Workbench is a 0-to-1 platform that consolidates the entire AI development lifecycle into a single, seamless environment. Designed for data scientists, ML engineers, and analysts, it replaced a fragmented ecosystem of 30+ tools and reduced average AI development time from 25 to 56 weeks down to 15 weeks.

Problem

AI development at scale was broken, and not because engineers lacked skill. The tools and knowledge they needed were scattered everywhere.

Data scientists, ML engineers, and analysts were navigating 30+ disparate tools to get their work done, stretching AI development timelines to anywhere from 25 to 56 weeks. But the deeper problem wasn’t the tools. It was what happened between them. Teams spent weeks gathering information through Slack threads, Zoom calls, and asking peers just to establish a course of action. Institutional knowledge lived in people’s heads, not in the systems where work happened.

I uncovered these pain points through follow-me-home research with data scientists, MLEs, and analysts, observing them in their actual work environments to understand where time and momentum were being lost.

Before

After

My Role

 

I served as lead designer on a 0-to-1 product, designing AI Workbench from the ground up with two supporting designers contributing to the experimentation and observability workflows. 

Research

I uncovered these pain points through follow-me-home research with data scientists, MLEs, and analysts, observing them in their actual work environments to understand where time was being lost. Of the three pain points identified, tribal knowledge and ineffective collaboration were felt most acutely by users.

My process also included journey mapping to understand the end-to-end AI development workflow, and competitive analysis drawing inspiration from platforms like Google Vertex, Kaggle, and Databricks to understand what patterns users might already be familiar with. Since no platform like this existed internally, every foundational design decision had to be built from scratch.

 

Ideation

After conducting competitive analysis inspiration from platforms like Google Vertex, Kaggle, and Databricks, I explored designs where the user can see the activity of their projects. 

A key early decision was the global navigation architecture. Rather than forcing a linear progression, I designed the left-hand nav to expose all workflow stages simultaneously, giving experienced users the freedom to move non-linearly among all ML objects while still providing a clear default path for newer users.

When I tested these design concepts with users, they got a better idea of what AI Workbench is solving but they did not know how to start navigating the product. 

I had to think about the starting point users were at, such as being onboarded to a completely new project, or finding data for the first time, etc.

On the project overview page, I thought about the multiple starting points users could take and made a card button for each step. I also added a call-to-action button to let the user create a new model asset because that’s the first artifact a user would have to create to start a ML project.

On the data page, I updated to an empty state and added a call-to-action button where the user can take the first action of exploring and finding data.

I used object-oriented UX framework to help me identify the actors, actions, and objects required to complete a task, which helped guide me in my designs.

In this data example, the user needs to explore data through Data Discovery and 

Exploration first, import it into AI Workbench

then over time, the user can keep track of data they have access to and can explore in 

notebooks. 

To help keep this project progressing without feeling like we were “boiling the ocean,” I scheduled weekly opmechs with engineering and product teams to align on scope, feasibility, and next actions. This would be the meeting where we dive deep into trade-offs between MVP vs target-state

Final Designs

AI Workbench consolidates the entire AI development lifecycle into a single, seamless platform. Instead of jumping between 30+ tools and reconstructing context each time, users move through one connected environment:

• • Projects. A shared workspace where teams collaborate around a common goal.

• • Data. Add and manage datasets within the project context.

• • Notebooks. Collaborative, real-time notebooks for data exploration where teammates can edit simultaneously and see each other’s changes live.

• • Pipelines. Build and manage training pipelines without leaving the platform.

• • Experiments. Run and compare model experiments in context.

• • Observability. Monitor model performance end-to-end.

The global left-hand navigation gives users access to every workflow stage at any time, because AI development isn’t always linear. Users can jump between stages freely without losing their place or their context.

Designing for Collaboration

Since tribal knowledge and ineffective collaboration were the most acutely felt pain points, collaboration wasn’t just a feature. It was a design principle that ran through the entire platform.

Every project creates a shared environment where team members see the same workflows, the same data, and the same state. Notebooks are collaborative by default, with real-time co-editing so users can see each other’s changes as they happen. A shared library of notebooks makes institutional knowledge visible and accessible to everyone. Things like how others have tackled similar problems or working examples of new concepts are no longer locked away with those who know who to ask.

Results

• • Average AI development time dropped from 25 to 56 weeks down to 15 weeks

• • 235 monthly active users on the alpha launch

Users felt the shift immediately:

“I used to spend half my day just setting up my environment. Now I can jump straight into building and iterating. It’s a game-changer.” — AI Scientist

“This library of notebooks is like a brainstorming partner. Seeing how others tackle problems and having a pre-built example to start from has helped me explore new concepts, like Agents and Tools, much faster than before.” — ML Engineer

“The sample notebooks were a perfect way to get up to speed. The seamless setup meant I could start contributing on day one without any frustration.” — New Team Member