Turning a fragmented model handoff into one continuous workflow.

Details

Role: Sole UX lead
Product: AI/ML model-as-a-service platform
Partners: Product, engineering, data science & domain experts
Scope: Workflow architecture, interaction design, testing & QA

Overview

BRYX is a browser-based machine-learning platform developed by KCI Technologies for engineering professionals. RoboClean removes noise from point cloud scans; RoboFlat uses cleaned point clouds to analyze concrete floor flatness.

RoboFlat model run with datasets and RoboClean results

Problem

One task. Two models. Too much work between them.

These models represented consecutive steps in one engineering task, but the interface treated them as separate destinations. Users had to move files between the two before they could continue their work.

Before / Manual handoff

A separate data-management task

  1. Locate and download the RoboClean output.
  2. Leave the model workflow and open dataset management.
  3. Upload the RoboClean output as a new dataset.
  4. Return to RoboFlat and locate the new RoboClean dataset.
  5. Select the dataset and complete validation.

After / Integrated handoff

Continue inside RoboFlat

  1. Start the RoboFlat run.
  2. Select an eligible RoboClean output alongside the other data options.
  3. Review dataset-creation feedback and complete validation.

BRYX creates the required dataset in the background and makes its storage consequence visible.

Solution

Translate the system requirement into a usable interface.

As the sole UX designer, I owned the workflow structure, component behavior, responsive states, system messaging, prototypes, usability-test updates, and implementation QA.

Challenge

Shorten the path. Keep the consequences clear.

The user’s intent was simple: use a cleaned output in the next model run. Three system constraints shaped the design.

Product & domain experts
Requirements, priorities, and operational rules.
Engineering & data science
Dataset creation, storage behavior, and implementation.
My role
Interaction model, interface hierarchy, states, and validation.
Create a dataset
RoboFlat could not consume the output until BRYX converted it into a platform dataset.
Respect account storage
The new dataset counted against the user’s available data allocation.
Fit a dense workflow
Run setup already had vertically stacked content and distant navigation controls.

Design decision / Data selection

Put the integration where the next decision happens.

Users needed RoboClean outputs to become a data option inside RoboFlat. Placing the integration at data selection connected the two models at the moment it mattered.

Before. Only previously created datasets appear.
Before. Only previously created datasets appear.
After. Eligible RoboClean outputs appear at the point of data selection.
After. Eligible RoboClean outputs appear at the point of data selection.

Design decision / System feedback

Make the invisible system action understandable.

The shortcut still created a dataset and consumed storage. The interface needed to explain that behavior and provide a recovery path when capacity was unavailable.

Capacity available: inform, then allow progress.

Informational feedback makes dataset creation and storage use explicit.
Informational feedback makes dataset creation and storage use explicit.

Capacity insufficient: block, then offer recovery.

A blocking state explains the storage constraint and offers a next action.
A blocking state explains the storage constraint and offers a next action.

Design decision / Responsive navigation

Keep progress and navigation in the same field of view.

The integration added pressure to a dense model-run step. I explored ways to consolidate navigation while preserving room for data, parameters, and validation feedback.

V1 / Rejected

Bottom sticky controls

Actions stayed visible, but a second sticky region compressed the usable viewport.

Exploration: persistent controls at the bottom of the workflow.
Exploration: persistent controls at the bottom of the workflow.

V2 / Refined

Controls beside the stepper

Progress and navigation came together, but the layout still needed to be responsive.

Exploration: navigation controls alongside workflow progress.
Exploration: navigation controls alongside workflow progress.

V3 / Selected direction

A responsive, consolidated pattern

Desktop keeps the workflow overview and navigation together. Narrow screens prioritize the current step, the next action, and enough context to keep users oriented.

V3 / Desktop

List component replaced table rows

Converting the Datasets and RoboClean Results tables to lists made the info more scannable and mobile friendly.

RoboFlat datasets in the selected desktop workflow Desktop. A shared location for progress and navigation.
Desktop. A shared location for progress and navigation.

V3 / mobile

Optimized for small screen size

Prioritizing the current step’s actions helped to determine what information remained prominent on small screens.

Annotations

  1. Model details hidden behind a tooltip.
  2. “Add Dataset” button label reduced to a plus icon.
  3. Metadata badges scroll horizontally.
  4. Progress indicator replaces numbered steps.
  5. “Previous” and “Next” button labels reduced to left and right chevron icons.
  6. List item details stacked vertically.
  7. Selection radio moved to the right side for improved right thumb access.

Framework

A reusable framework for the next model.

RoboFlat shared a recognizable operating structure with other BRYX models while retaining its own configuration and validation needs. The integration became part of a broader model-run framework.

Shared model run framework
Shared model run framework
RoboFlat. The shared workflow structure with model-specific requirements.
RoboFlat. The shared workflow structure with model-specific requirements.
RoboClean. The same stage within a consistent operating structure.
RoboClean. The same stage within a consistent operating structure.

Usability validation

Follow-up testing showed that the integration was easier to find and complete.

In a follow-up usability test with five participants, users reached the integration through RoboFlat and completed the workflow in fewer steps than the previous manual handoff.

Original and updated task flows: from an available RoboClean output to selected, validated data in RoboFlat.
Original and updated task flows: from an available RoboClean output to selected, validated data in RoboFlat.
Updated workflow: connect RoboClean results directly inside RoboFlat
Updated workflow: connect RoboClean results directly inside RoboFlat

Observed change

The handoff became part of the model workflow instead of a separate data-management task.

Implementation & QA

The design work continued into implementation.

I reviewed the implemented experience in the test environment and documented discrepancies through Figma comments and the team’s QA workflow. Working with product and engineering, I carried the interaction details through review, correction, and verification. This loop connected the design intent to the implemented experience.

Outcome & reflection

A shorter workflow that still told users what the system was doing.

The tested experience connected RoboClean and RoboFlat at data selection, made dataset creation and storage consequences visible, and improved navigation through a dense model-run workflow.

Complex product design is often about deciding when users need to understand the system’s rules. Here, the right answer was to hide the manual handoff while making its consequences clear.