PROCESSING TIME
~-50%
PROJECT - 01
Product Design · Trust & Safety
Queue Zero
Designing the human safety net in an AI pipeline
ROLE
Solo · E2E
TEAM
PO · PM · Dev
DURATION
~5 months
DELIVERABLE
Deployed
TYPE
Human-AI Interface
AI detects media requiring action, but it sometimes makes mistakes. Misclassification leads to either unjustified removal or online violations.
PROCESSING TIME
~-50%
PROJECT - 01
Product Design · Trust & Safety
Queue Zero
Designing the human safety net in an AI pipeline
ROLE
Solo · E2E
TEAM
PO · PM · Dev
DURATION
~5 months
DELIVERABLE
Deployed
TYPE
Human-AI Interface
AI classifies inspection images and occasionally makes mistakes. An undetected error leads to a contractual dispute for the service provider.
PROJECT - 01
Product Design · Trust & Safety
Queue Zero
Designing the human safety net in an AI pipeline
ROLE
Solo · E2E
TEAM
PO · PM · Dev
DURATION
~5 months
TYPE
Human-AI Interface
DELIVERABLE
Deployed
AI detects media requiring action, but it sometimes makes mistakes. Misclassification leads to either unjustified removal or online violations.
PROCESSING TIME
~-50%
PROJECT - 01
Product Design · Trust & Safety
Queue Zero
Designing the human safety net in an AI pipeline
ROLE
Solo · E2E
TEAM
PO · PM · Dev
DURATION
~5 months
DELIVERABLE
Deployed
TYPE
Human-AI Interface
AI detects media requiring action, but it sometimes makes mistakes. Misclassification leads to either unjustified removal or online violations.
Results
Results
"This mode makes everything much more readable. The path is clearly marked out, with no need to guess what comes next."
Expert moderator, prototype V2 test
A team member saw Queue Zero as a direct answer to what they dealt with every day. A signal from the inside.
18 story pages, 1 to 2 developer clarifications. The expert moderator (the one who feared being constrained) reported zero friction during the retest.
"This mode makes everything much more readable. The path is clearly marked out, with no need to guess what comes next."
Expert moderator, prototype V2 test
A team member saw Queue Zero as a direct answer to what they dealt with every day. A signal from the inside.
18 story pages, 1 to 2 developer clarifications. The expert moderator (the one who feared being constrained) reported zero friction during the retest.
"This mode makes everything much more readable. The path is clearly marked out, with no need to guess what comes next."
Expert moderator, prototype V2 test
A team member saw Queue Zero as a direct answer to what they dealt with every day. A signal from the inside.
18 story pages, 1 to 2 developer clarifications. The expert moderator (the one who feared being constrained) reported zero friction during the retest.
~-50%
Estimated processing time
Structured navigation, prioritized processing, time-saving
Based on field feedback, there is a 30% to 45% gain. While the design portion cannot be isolated, qualitative indicators confirm these results. Data has been adjusted for confidentiality.
~-50%
Estimated processing time
Structured navigation, sequential processing; saves time
Based on field feedback, there is a 30% to 45% gain. While the design portion cannot be isolated, qualitative indicators confirm these results. Data has been adjusted for confidentiality.
~-50%
Estimated processing time
Structured navigation, prioritized processing, time-saving
Based on field feedback, there is a 30% to 45% gain. While the design portion cannot be isolated, qualitative indicators confirm these results. Data has been adjusted for confidentiality.
Business context
Business context
From reported content to sanction
A user-generated content platform receives thousands of reported or automatically detected media assets every day. An AI model classifies each asset by violation type and severity level, triggering removal, account suspension, or legal escalation. Before taking action, the moderation team reviews each asset in a thumbnail gallery, a step that AI cannot complete on its own.
ROV hull inspection: from drone to report.
A user-generated content platform receives thousands of reported or automatically detected media assets every day. An AI model classifies each asset by violation type and severity level, triggering removal, account suspension, or legal escalation. Before taking action, the moderation team reviews each asset in a thumbnail gallery, a step that AI cannot complete on its own.
From reported content to sanction
A user-generated content platform receives thousands of reported or automatically detected media assets every day. An AI model classifies each asset by violation type and severity level, triggering removal, account suspension, or legal escalation. Before taking action, the moderation team reviews each asset in a thumbnail gallery, a step that AI cannot complete on its own.
Issue
Issue
"A gallery without order, without landmarks. It was impossible to know where I was in the batch."
What the moderator was experiencing
The moderator arrived at a gallery of raw thumbnails containing a mix of photos and videos, with no clear sequence, no distinction between approved and pending items, and no progress indicators. The AI errors was concentrated on ambiguous cases, such as video thumbnails lacking real context. Correcting a misclassified media item required search blindly through the gallery.
Three hours of onboarding, a ten-page guide, forgotten in two weeks. The tool should have driven the process, not just detailed it.
"A gallery without order, without landmarks. It was impossible to know where I was in the batch."
What the moderator was experiencing
The moderator arrived at a gallery of raw thumbnails containing a mix of photos and videos, with no clear sequence, no distinction between approved and pending items, and no progress indicators. The AI errors was concentrated on ambiguous cases, such as video thumbnails lacking real context. Correcting a misclassified media item required search blindly through the gallery.
Three hours of onboarding, a ten-page guide, forgotten in two weeks. The tool should have driven the process, not just detailed it.
"A gallery without order, without landmarks. It was impossible to know where I was in the batch."
What the moderator was experiencing
The analyst arrived at a raw gallery: no sequences, no distinction between verified and untreated, no indicators. Errors were concentrated on boundary conditions (micro-fouling, degraded coating). Correcting a misclassified image meant searching blindly.
Three hours of onboarding, a ten-page guide, forgotten in two weeks. The tool should have driven the process, not just detailed it.
Approach
Approach
First project. Zero established processes.
Without a design peer, every assumption was documented in writing: justifiable decisions, defensible pivots. A senior UX designer was available occasionally to challenge roadblocks. The domain was mastered in two months: with no internal documentation, information was gathered directly from the team. A limited design system, adapted component by component. Shape Up methodology: the scope is negotiated, not the deadline.
First project. No established process.
Without a design peer, every assumption was documented in writing: justifiable decisions, defensible pivots. A senior UX designer was available occasionally to challenge roadblocks. The domain was mastered in two months: with no internal documentation, information was gathered directly from the team. A limited design system, adapted component by component. Shape Up methodology: the scope is negotiated, not the deadline.
First project. Zero established processes.
Without a design peer, every assumption was documented in writing: justifiable decisions, defensible pivots. A senior UX designer was available occasionally to challenge roadblocks. The domain was mastered in two months: with no internal documentation, information was gathered directly from the team. A limited design system, adapted component by component. Shape Up methodology: the scope is negotiated, not the deadline.
01
Immersion
Documentary review. Observation of a live moderation session. Journey maps from session recordings and gallery usage. Heuristic audit (Nielsen grid, 6 categories). First prototype submitted to the team in a workshop. The project paused to allow for an update of the classification model.
The model sorts by severity level. The moderator needs to see the category immediately to know which action to apply. The immersion makes the gap visible.
01
Immersion
Documentary review. Observation of a live moderation session. Journey maps from session recordings and gallery usage. Heuristic audit (Nielsen grid, 6 categories). First prototype submitted to the team in a workshop. The project paused to allow for an update of the classification model.
The model sorts by severity level. The moderator needs to see the category immediately to know which action to apply. The immersion makes the gap visible.
01
Immersion
Documentary review. Observation of a live moderation session. Journey maps from session recordings and gallery usage. Heuristic audit (Nielsen grid, 6 categories). First prototype submitted to the team in a workshop. The project paused to allow for an update of the classification model.
The model sorts by severity level. The moderator needs to see the category immediately to know which action to apply. The immersion makes the gap visible.
02
Tests
Three sessions, three profiles: beginner, intermediate, and expert moderators, recruited internally, who already handle similar cases daily. A second formal validation workshop for the Severity pivot (PO, PM, tested users, UX).
Four parties, one consensus. The Severity pivot was no longer just a design decision.
02
Tests
Three sessions, three profiles: beginner, intermediate, and expert moderators, recruited internally, who already handle similar cases daily. A second formal validation workshop for the Severity pivot (PO, PM, tested users, UX).
Four parties, one consensus. The Severity pivot was no longer just a design decision.
02
Tests
Three sessions, three profiles: beginner, intermediate, and expert moderators, recruited internally, who already handle similar cases daily. A second formal validation workshop for the Severity pivot (PO, PM, tested users, UX).
Four parties, one consensus. The Severity pivot was no longer just a design decision.
03
Handoff
Multi-profile decision matrix, mockup iterations, and writing User Stories (10 US, Given/When/Then format with acceptance criteria, typically written by the PM).
18 pages in Given/When/Then format, which remained the technical reference until the end of the build phase
03
Handoff
Multi-profile decision matrix, mockup iterations, and writing User Stories (10 US, Given/When/Then format with acceptance criteria, typically written by the PM).
18 pages in Given/When/Then format, which remained the technical reference until the end of the build phase
03
Handoff
Multi-profile decision matrix, mockup iterations, and writing User Stories (10 US, Given/When/Then format with acceptance criteria, typically written by the PM).
18 pages in Given/When/Then format, which remained the technical reference until the end of the build phase
Decisions
Decisions
Severity/Category Pivot
The V1 design organized the gallery by severity level, inherited directly from the model output: the violation category (Spam, Harassment, Misinformation, etc.) was relegated to a secondary filter, invisible without clicking.
Decision: Category in navigation, severity omnipresent. Moderators need the category to act fairly, but severity must remain visible everywhere.
Severity/Category Pivot
The V1 design organized the gallery by severity level, inherited directly from the model output: the violation category (Spam, Harassment, Misinformation, etc.) was relegated to a secondary filter, invisible without clicking.
Decision: Category in navigation, severity omnipresent. Moderators need the category to act fairly, but severity must remain visible everywhere.
01
AI Model
Sidebar V1
Navigation by severity level, structured based on classifier output: offense category is hidden until clicked.
01
AI Model
Sidebar V1
Navigation by severity level, structured based on classifier output: offense category is hidden until clicked.
01
AI Model
Sidebar V1
Navigation by severity level, structured based on classifier output: offense category is hidden until clicked.
01
AI Model
Sidebar V1
Navigation by severity level, structured based on classifier output: offense category is hidden until clicked.
02
Maintenance signal
Field insights
The moderator reveals: severity was already guiding the navigation, but without the visible category, they never knew which action to apply.
02
Maintenance signal
Field insights
The moderator reveals: severity was already guiding the navigation, but without the visible category, they never knew which action to apply.
02
Maintenance signal
Field insights
The moderator reveals: severity was already guiding the navigation, but without the visible category, they never knew which action to apply.
02
Maintenance signal
Field insights
The moderator reveals: severity was already guiding the navigation, but without the visible category, they never knew which action to apply.
03
Severity Pivot
Sidebar V2
First-level navigation category (aggregated severity badge), sub-category stepper sorted by descending severity. Cost: V1 to be undone. Non-optional.
03
Severity Pivot
Sidebar V2
First-level navigation category (aggregated severity badge), sub-category stepper sorted by descending severity. Cost: V1 to be undone. Non-optional.
03
Severity Pivot
Sidebar V2
First-level navigation category (aggregated severity badge), sub-category stepper sorted by descending severity. Cost: V1 to be undone. Non-optional.
03
Severity Pivot
Sidebar V2
First-level navigation category (aggregated severity badge), sub-category stepper sorted by descending severity. Cost: V1 to be undone. Non-optional.
The decision bar: act without leaving
In V1, there were two separate actions: qualifying on a dedicated page, then penalizing via a hard-to-read menu within a gallery of thousands of thumbnails of all media types. To consolidate without overloading, a permanent sidebar and a full-screen modal panel were ruled out through testing.
Decision: The decision bar appears at the bottom of the gallery only when requested. Collapsed by default, it allows the thumbnail grid to maintain its full display width.
The decision bar: act without leaving
In V1, there were two separate actions: qualifying on a dedicated page, then penalizing via a hard-to-read menu within a gallery of thousands of thumbnails of all media types. To consolidate without overloading, a permanent sidebar and a full-screen modal panel were ruled out through testing.
Decision: The decision bar appears at the bottom of the gallery only when requested. Collapsed by default, it allows the thumbnail grid to maintain its full display width.
Feature
Decision bar
A module that remains hidden out of sight until needed. Collapsed by default, it never interferes with the gallery.
Bottom anchored
Closed by default
Feature
Decision bar
A module that remains hidden out of sight until needed. Collapsed by default, it never interferes with the gallery.
Bottom anchored
Closed by default
OPTIONS CONSIDERED AND REJECTED
- Permanent sidebar: confirmed cognitive overload - Full-screen modal panel: breaks the flow in gallery view - Tabs: too much navigation that could lead to user error
OPTIONS CONSIDERED AND REJECTED
- Permanent sidebar: confirmed cognitive overload - Full-screen modal panel: breaks the flow in gallery view - Tabs: too much navigation that could lead to user error
Isolated Component
Decision bar open
Multi-selection of media, application of a common sanction. Removal, account suspension, or legal escalation.
Isolated Component
Decision bar open
Multi-selection of media, application of a common sanction. Removal, account suspension, or legal escalation.
STAGE FLOW
1
Select: click a thumbnail, or select multiple
2
Qualify: Decision bar open (Type / Severity / Context)
3
Apply: decision confirmed
STAGE FLOW
1
Select: click a thumbnail, or select multiple
2
Qualify: Decision bar open (Type / Severity / Context)
3
Apply: decision confirmed
Prevention rather than correction
Developing a cancel button represented a technical cost disproportionate to the expected benefit, which was out of scope for the current Shape Up cycle. The real question was not how to cancel, but why this need existed: the error remains invisible when it occurs, the same blind spot that existed at the start of the project.
Outcome: zero media left behind in the gallery. The system guides toward completion, not cancellation.
Prevention rather than correction
Developing a cancel button represented a technical cost disproportionate to the expected benefit, which was out of scope for the current Shape Up cycle. The real question was not how to cancel, but why this need existed: the error remains invisible when it occurs, the same blind spot that existed at the start of the project.
Outcome: zero media left behind in the gallery. The system guides toward completion, not cancellation.
Feature
Forgetfulness alerts
Two mechanisms: a severity badge on the category, a badge on the relevant subcategory, and a summary modal at the end of the process.
Preventive
Real-time
Feature
Forgetfulness alerts
Two mechanisms: a severity badge on the category, a badge on the relevant subcategory, and a summary modal at the end of the process.
Preventive
Real-time
OPTIONS CONSIDERED AND REJECTED
- Undo button: technically complex for marginal gain (Shape Up) - Decision: address the root cause (visibility) rather than the symptom (cancellation)
OPTIONS CONSIDERED AND REJECTED
- Undo button: technically complex for marginal gain (Shape Up) - Decision: address the root cause (visibility) rather than the symptom (cancellation)
Isolated Component
Missed Alerts Modal
Triggered if media remains unverified in a subcategory.
Isolated Component
Missed Alerts Modal
Triggered if media remains unverified in a subcategory.
STAGE FLOW
1
Processing: severity badges visible in real time
2
Navigation: the moderator leaves a category, badges indicate unverified subcategories
3
End of journey: modal triggered, list of relevant subcategories
STAGE FLOW
1
Processing: severity badges visible in real time
2
Navigation: the moderator leaves a category, badges indicate unverified subcategories
3
End of journey: modal triggered, list of relevant subcategories
Filters: consistency and readability
Filters were used during every session. The modal opened below the button, but it was too small, making scrolling mandatory. There was no hierarchy or color coding, even though the severity chips had already established this visual language elsewhere in the interface. This created a break in an interface that had otherwise built its own cohesive visual vocabulary.
The result is that no single dimension remains isolated. The filters now cover the cross-references that the moderator was already making mentally.
Filters: consistency and readability
Filters were used during every session. The modal opened below the button, but it was too small, making scrolling mandatory. There was no hierarchy or color coding, even though the severity chips had already established this visual language elsewhere in the interface. This created a break in an interface that had otherwise built its own cohesive visual vocabulary.
The result is that no single dimension remains isolated. The filters now cover the cross-references that the moderator was already making mentally.
Feature
Redesigned filters
Visual hierarchy and color codes aligned with severity badges: immediate recognition, no relearning required.
Color codes
Visual Hierarchy
Feature
Redesigned filters
Visual hierarchy and color codes aligned with severity badges: immediate recognition, no relearning required.
Color codes
Visual Hierarchy
Feature
Redesigned filters
Visual hierarchy and color codes aligned with severity badges: immediate recognition, no relearning required.
Color codes
Visual Hierarchy
Feature
Redesigned filters
Visual hierarchy and color codes aligned with severity badges: immediate recognition, no relearning required.
Color codes
Visual Hierarchy
BEFORE THE REDESIGN
- Tiny modal below the button, scrolling required - No color coding: deviation from the existing visual language - Raw text in an excessively small space - No design system applied
BEFORE THE REDESIGN
- Tiny modal below the button, scrolling required - No color coding: deviation from the existing visual language - Raw text in an excessively small space - No design system applied
Isolated Component
Sidebar filters
Right sidebar, overlaying the decision sidebar, with filters organized by category for a readable hierarchy at a glance. An extension of the interface, not a disruption.
Isolated Component
Sidebar filters
Right sidebar, overlaying the decision sidebar, with filters organized by category for a readable hierarchy at a glance. An extension of the interface, not a disruption.
STAGE FLOW
1
Open: filters button
2
Select: filters with colored severity chips
3
Apply: updated gallery, active filters visible
STAGE FLOW
1
Open: filters button
2
Select: filters with colored severity chips
3
Apply: updated gallery, active filters visible
The correction surface
The right sidebar is the only place where severity is actually verified and corrected—re-categorized if needed—using the filter color codes. Correction is done there, not in the gallery.
Even in borderline cases, the severity level remains identifiable at a glance. You won't make corrections that go against your intent.
The correction surface
The right sidebar is the only place where severity is actually verified and corrected—re-categorized if needed—using the filter color codes. Correction is done there, not in the gallery.
Even in borderline cases, the severity level remains identifiable at a glance. You won't make corrections that go against your intent.
Feature
Correction sidebar
Simple click on a media item: color-coded severity chips and category reassignment, without leaving the gallery.
Correction
Classification
Feature
Correction sidebar
Simple click on a media item: color-coded severity chips and category reassignment, without leaving the gallery.
Correction
Classification
BEFORE THE REDESIGN
- No-code color dropdown, only text to browse through - No indication of the action to take
BEFORE THE REDESIGN
- No-code color dropdown, only text to browse through - No indication of the action to take
Isolated Component
Severity badges and reassignment
Color states (Low / Moderate / High / Critical), following the same visual logic as the filters. Reassignment to the correct category.
Isolated Component
Severity badges and reassignment
Color states (Low / Moderate / High / Critical), following the same visual logic as the filters. Reassignment to the correct category.
STAGE FLOW
1
Select media to process from the gallery
2
Verify: severity chips, color codes, correction if needed
3
Reassign: if the media belongs to another category
STAGE FLOW
1
Select media to process from the gallery
2
Verify: severity chips, color codes, correction if needed
3
Reassign: if the media belongs to another category
What was delivered
Designed from scratch: Decision bar (bulk selection and common action), navigation by category and omnipresent severity, subcategory stepper, Forgotten Alerts modal, Help modal (support + guides + shortcuts).
Redesigned based on the existing version: correction sidebar, redesigned filters, thumbnail gallery, page header, and bookmark icon (supporting evidence).
What was delivered
Designed from scratch: Decision bar (bulk selection and common action), navigation by category and omnipresent severity, subcategory stepper, Forgotten Alerts modal, Help modal (support + guides + shortcuts).
Redesigned based on the existing version: correction sidebar, redesigned filters, thumbnail gallery, page header, and bookmark icon (supporting evidence).
Reflection
Reflection
What I would have done differently
With direct access to users right from the discovery phase, the Severity pivot would have emerged before testing rather than after. A more mature design system would have allowed visual hierarchy alternatives to be resolved with concrete support, rather than based on judgment alone. Internal recruitment remained the only realistic option in this context: this impacts the external validity of the feedback obtained, a compromise that must be acknowledged rather than ignored.
First project in this field, without a peer designer for continuous review. It isn't about adapting to ambiguity: it is about building the method as we move forward.
What I would have done differently
With direct access to users right from the discovery phase, the Severity pivot would have emerged before testing rather than after. A more mature design system would have allowed visual hierarchy alternatives to be resolved with concrete support, rather than based on judgment alone. Internal recruitment remained the only realistic option in this context: this impacts the external validity of the feedback obtained, a compromise that must be acknowledged rather than ignored.
First project in this field, without a peer designer for continuous review. It isn't about adapting to ambiguity: it is about building the method as we move forward.