Reimagining Compliance Review for Retail Operations

The approach

Making room for the work

How do we create more space for cards and new filter options?

In the original gallery, the filter panel claimed a large share of the page’s width at the expense of the image cards. Expanding it reduced the layout to a three-card grid, limiting how many images users could review at once. As more filters were added, the panel also became increasingly difficult to scale.

To give the image cards as much horizontal space as possible, I moved the filters into a floating panel that stayed within reach as users scrolled. I paired this with infinite scrolling instead of pagination, so users could browse submissions without having to repeatedly switch between pages. I also reorganized the filter panel into two columns, making room for additional filters as the feature grew.

Automation with a human backstop

How do we keep the human in the loop while the new AI is still learning?

Store employees sometimes submitted images that weren’t relevant for compliance reviews, such as photos of floors, unrelated subjects, or selfies. One Door’s in house image-recognition model was still learning to detect irrelevant images correctly, so its results weren’t always accurate.

I took the idea of ‘flagging’ literally, and used flag icons to mark images the model identified as potentially irrelevant. These 'flags' made it easier for HQ reviewers to quickly filter and assess non-compliant images. Reviewers could also remove incorrect flags, which provided feedback that helped the image recognition model learn from its mistakes. The model acted as a first-pass filter, not a decision-maker.

Every image has a history

How do we make sense of images uploaded at any time and in any sequence?

Previously, resubmitted images appeared as regular image submission records. Reviewers had to sift through hundreds (sometimes thousands) of images to find the latest resubmission and check whether a previously reported issue had been fixed.

I updated the card labels to highligth image states: 'Submitted', 'Resubmitted', 'Revoked'. I also introduced a timeline in the image detail view. This linked the original submission to its resubmissions, which brought related images, statuses, dates, and comments into one view. Reviewers could now quickly find the relevant images, compare submissions, and verify whether the store had addressed the issue without having to search or filter through the entire gallery.

Closing the loop on compliance

How do we keep the entire compliance workflow in the app?

When HQ rejected an image, store employees needed to understand why, make corrections, and resubmit it. Historically, communication for rejected images was scattered across emails, internal team messages/calls, and other channels, making it harder to track issues and follow-up.

I brought these steps closer together by designing a component in the store-facing app that explained why an image was revoked and enabled direct communication between stores and HQ. I also introduced in-app notifications that alerted store employees to revoked images and HQ reviewers to new or resubmitted images awaiting review.

Later down the road, I also designed and introduced a missing-images report to help HQ identify required images that stores had not yet submitted. Together, these changes helped HQ identify missing submissions and communicate the corrections stores needed to make. Reviewers could then assess updated images within the same product. Both teams had a clearer view of what needed attention and what to do next.

Bringing the evolving experience together

How do we modernize the AI-assisted image status experience?

After a few months on other projects, I returned to Store Gallery to find that other teams had updated it with a new image engine, new workflows, and additional functionality. At the same time, other One Door products were adopting a more consistent visual language and interaction patterns. My goal was to bring Store Gallery in line with these broader changes.

I moved filters into a side panel. Image-specific actions now appeared in a bottom bar, but only when reviewers selected images. This kept the interface focused until those actions were needed. Secondary actions moved into an overflow menu to reduce clutter. A new attribute selector let reviewers choose which image attributes to display. The individual image view was also updated to match newer patterns.