Back to selected work
JSTOR DIGITAL STEWARDSHIP SERVICES • PRODUCT DESIGN
Designing a clearer path to digital stewardship
TIMELINE
1 Year
ROLE
Product Designer
TEAM
Design, product, engineering
Summary
JSTOR Digital Stewardship Services is a new AI-assisted platform that helps librarians and archivists process, describe, manage, and share digitized content. Institutions reported up to 95% of their collections sitting undigitized or undescribed due to understaffing, budget cuts, and time constraints. I owned 5 aspects of the platform, designing for 120+ library leaders.
90% reduction reported
of hands-on processing
86% reduction of error tickets
for the Support team
2 institutions fully migrated
their content from self-hosted platform within 1 month
318+ institutions
on the platform after launch
OVERVIEW
Helping cultural institutions make more of their collections discoverable.
JSTOR Digital Stewardship Services (JDSS) is a platform that uses a new AI processing tool, Seeklight, to help librarians and archivists generate descriptions, transcriptions, and enhance and enrich their metadata at both the item and collection levels. This allows them to increase the discoverability of their digitized content across the world, all without compromising human judgment..
DEFINING
What do we know
We entered the project with three working assumptions about Tier 3 institutions: they pay for Seeklight, so they should be able to choose field by field where it applies; Seeklight-enablement should therefore be the default in metadata templates for new projects; and these are digital management-first teams reaching for a powerful AI tool. Concept testing with internal users confirmed the direction and reshaped the mapping experience.
01
02
03
04
05
Outcome: Start small, use simple terminology, make mapping safe to edit and undo.
What do we want to know
The unknowns were sorted into five buckets. By grouping them together with engineering, data and product, it was made clear which answers were blocking design and which could wait for a later phase of release.
USER RESEARCH QUESTIONS
Minimum mapping functionality users will accept?
Does a mixed tags column confuse when only some fields are generated?
Do un-mappable fields like Linked Authorities kill adoption?
TECHNICAL FEASIBILITY QUESTIONS
What are our known tech limitations?
Will OAI be impacted by schema change?
Does project size impact metadata and summary generation time?
Impact of filtering? FE dynamically builds the table…if tags/no tags are there, will we have empty table cells?
PRODUCT QUESTIONS
How will we treat system generated tags (AI, edited, reviewed?) Does this pose buisness implications in our cost model?
Is publishing-target mapping the same experience as Seeklight mapping?
Do non-metadata Seeklight features enable for Tier 3
BUSINESS RULES & DATA QUESTIONS
Tier 3 -> Tier 2 downgrades and whether they're in scope
More tiers next? Or just bundled features?
Collisions where "reviewed" may already mean something
Participant targets and support for large projects
USER INTERVIEWS & OUTCOMES
Uncovering the practical barriers behind unprocessed collections.
Through conversations with more than 120 library leaders, I learned that constrained staff, low funding, and barely enough time left teams holding expanding collections with limited infrastructure. For some, as much as 95% of their content remained undigitized, undescribed, or underdescribed.
Discovery
Unprocessed materials
Outdated platform
Cumbersome to maintain
Limited resources
Staff, funding, and time
As collections grew, the capacity to process them was outpaced, leaving up to 95% of unprocessed material on the shelves and without descriptions for content, their items were at risk of being undiscoverable by the general public.
“Backlogs are not just a logistical issue but also an emotional burden for many staff, who see unprocessed materials as a mark against their professional and ethical responsibilities.”
SPECIAL COLLECTIONS LIBRARIAN, REVEAL DIGITAL
Key findings
Earlier, I had assumed that all users who subscribed to Seeklight would want their content to go through it with each upload. That was found to be false after research. Users actually wanted the ability to turn on and turn off the functionalities they were subscribed to. That gave me a new insight to incorporate into the design.
After unpacking and sorting, I was able to parse out what key trends popped up throughout the conversations; things that worked on old cataloguing platforms and what would be beneficial to keep, enhance, or dispose of.
Control & Trust
Administrators need to be able to edit and approve any media items that go through Seeklight. This means the system has to rate its confidence level to signal potential error or hallucinations during processing.
Ease of use
A wide variety of user types (student workers, veteran cataloguers, librarians, part-time volunteers) need to be able to easily use the site. The experience needs to be unambiguous.
Large volume control
Hundreds of items will be uploaded to the JDSS platform, which means seamless management is paramount to the experience.
Improving what I know
As we built out the new JDSS experience, one aspect I wanted to focus on was user relief surrounding the Activity Monitor. The current Activity Monitor only showed that an item had failed to upload, but not why the item had failed to upload.
When my request to focus on this was deprioritized for several sprints, I reached out to the Outreach team and requested concrete user data.
27
logged tickets for “unclear error messaging,” our biggest user issue in 2024/2025
So, I led a design and development workshop around how we can leverage our current data into a comprehensive alert system through Frontend and Backend information:
Open questions
Activity monitor should:
Outcome: Incremental implementation. We started with project-to-project upload errors.
DESIGN PROCESS
Designing for familiar workflows and new possibilities.
Before moving into the detailed interface work, I defined a high-level sitemap. The primary goal was to organize and structure all the content from the research phase and the past platform's experience. The goal was to make the advanced capabilities of AI and its complementary features feel approachable rather than disruptive.
Adding media within a project
Make groups, describe content, create transcripts, re-run failed uploads
Upload flow
Main
Start journey
Statistics
Usage
Create, access usage, and manage projects.
Homepage flow
Designs.
The new experience made it easier for teams to move through uploading, grouping, reviewing, and publishing content while keeping the human decision-maker in control. This work also reinforced that clear states, meaningful error messages, and practical documentation are just as important as polished screens.
01
Here are the core pathways: Create, manage, or generate with Seeklight. The purpose of such prominent designing was so brand-new users are never left guessing where on where to start. In the same vein of "directing the user", the Recently Accessed Projects was pulled to the front page so that different users could pick up right where a colleague left off; one of the key valuable pieces noted in interviews.
02
EASE OF USE
Walkthrough length · 0:14
Due to so much content, institutions want to be aware of what their statistics look like in order to make informed upload decisions. For example, with audio and video using up purchased Seeklight tokens, librarians may make the decision to upload a transcript (text) of a speech instead of the 1-hour-long commencement address (video), thus saving Seeklight usage.
The expanded content cards catered to the user interviews where they pushed for transparency: Audio and video show hours used, while AI items show item counts, while plainly stating that deleting processed content frees storage but won't lower the item count.
03
CONTROL & TRUST
LARGE VOLUME CONTROL
Walkthrough length · 0:22
In past platforms, users never got a preview of the schema they were picking, not to mention how they would map to JSTORs new Seeklight Schema. Users pick a schema, new or duplicated from an existing project, before they name the project and move straight into mapping. One of my key assumptions was corrected here. I assumed that institutions who subscribed to Seeklight would want to use it…turns out they don't always want to. We'd need two separate flows. To make it easier on them, I added a toggle feature per project. It became a single, critical decision point on this page mapping page.
Mapping AI fields
Walkthrough length · 0:19
Here is the decision to opt in or out. If they opt in, users are exposed to the full list of Seeklight fields that will match to their project fields. Once accepted, to honor clarity, accordions and instructions on functionality and definitions were littered throughout the mapping experience, making the journey easier to digest.
04
CONTROL & TRUST
Walkthrough length · 0:25
Users can upload media file-by-file or by whole folder, while also being able to remove singular files without having to restart the upload flow. By selecting 'Group by Filename', files with matching filename syntax are identified and sequenced to associate the files with the same record, and turn them into a compound object: 3 pages from a 1900s book become one record, two scans of a postcard become another; groups can still be broken apart or deleted. Before the generation runs, users can add any extra information in the optional text box! It gives Seeklight some extra info and keeps humans part of the experience.
05
LARGE VOLUME CONTROL
Walkthrough length · 0:23
Bulk edit used to be its own page, with its own load time and backend calls. I made the decision to change the experience to a full-width sheet, which allowed cut those calls entirely, which in turn reduced loading time for users. They no longer had to be kept waiting, and the new design kept them in one continuous workflow.
Bulk edit does two jobs: it confirms fields that already match and also flags the ones that don't with a "Multivalue" marker. Users can either type in their correction or hit "Paste from Main", which pushes the starred record's metadata across the whole selection of items. By using that button, cataloguers can align like-items in the fastest way possible. In addition, every change leaves a trail: an Edited tag joins the AI-Generated tag, and microtext inside each column marks which fields have been edited by a human!
Less backend calls
06
EASE OF USE
CONTROL & TRUST
Walkthrough length · 0:13
The result of my error workshop.
After working with the Support and Engineering teams, I was able to create a more transparent error experience. Support reported less error inquiries once we launched the new experience which allowed users to pinpoint exactly what went wrong; even down to which item in a compound object was the problem. By surfacing those error types, I was able to turn a support bottleneck into something users could resolve themselves.
Less error inquiries
07
CONTROL & TRUST
Walkthrough length · 0:26
Linked Authority used to be inaccessible, meaning volunteers had to be able-bodied to help with cataloguing at all. I redesigned it by fixing the accessibility gaps and improving visual hierarchy. I worked closely with metadata librarians to get their needs met and made sure that my aim for simplicity didn't accidentally result in the feature losing dexterity.
Here, you can also see the flow of marking an AI-Generated item as reviewed. A "Reviewed" tag is added and lets others know it is ready to be published.
RESULTS
What changed
90%
reduction in hands-on time for an AI-assisted metadata project
West Virginia University
500+
photographs processed in under 45 minutes
Rollins College
2
institutions retired self-hosted systems for JSTOR Digital Stewardship Services
Vanderbilt and UDC
What to look for
next.
With JDSS being a completely new offering, majority feedback was qualitative. There was not enough user data to compare platform usage activity. However, as time passes, the data I anticipate being most useful for iterative discussions will be.
01
Time spent
02
Drop-off rates
03
Support ticket content
Want to talk more? Contact me!