Schema Builder (diff + approval)
- Governed changes: Blast radius checks, policy lint, two-person approval.
- Safe rollback: Versioned model with “before/after” receipts.
CASE STUDY • ENTERPRISE DATA • PUBLISHING PLATFORM
Metadata Management & Publishing:
One system for schema, pricing, production & compliant releases.
I led the UX design of a multi-module platform that replaces a legacy vendor tool with a cohesive, explainable workflow across Publishing, Production, and Pricing optimized for safety, speed, and scale.
System thinking
Schema Builder
AI Search (chat → chips)
Metadata Insights Assistant
Feed Console
- One object model shared across modules no more tool-hopping.
- Explain first, then act every change previews impact and has undo.
- Search that composes natural language → typed chips → saved views.
Context: One platform that changes safely and shows impact
We combined title, pricing, production, and feeds into one model with guardrails and receipts.
Large publishers behave like distributed systems: formats, territories, rights, partners, and price windows evolve on different clocks. The old stack spread this work across vendors and tools; schema edits meant tickets; downstream impact was discovered late—often by partners.
We redesigned it as a single platform where policy is visible, changes are previewable, and history writes itself. This required three big bets: a UI-native metadata builder (no vendor queue), a Metadata Insights Assistant to surface explainable fixes, and an AI Search that turns plain language into precise, reusable filters.
Click to zoom
Problem: Risk felt higher than return
Work spanned apps, rules were implicit, and impact was invisible.
Click to zoom
Cognitive overloadLegal, pricing, rights, partners scattered across tools with different rules.
Trust gapsNew users couldn’t evaluate consequences or fix issues safely.
Operational dragSchema edits and approvals took weeks; onboarding took too long.
Hidden breakageFeed failures surfaced late; recovery paths weren’t clear.
Solution: Clarity first. Control always. Compliance in view.
Two flagship capabilities, then supporting pillars.
Flagship capability
Metadata Insights Assistant
Click to zoom
Click to zoom
What this proves
The system can propose safe changes without acting like a black box. Suggestions rank by confidence and cite their basis. Users preview a diff, apply, and get a receipt. If it’s wrong, undo restores state and history remains intact.
Flagship capability
AI Search (chat → chips)
Click to zoom
Click to zoom
What this proves
What this proves: Anyone can ask in plain English; the system converts intent into typed chips you can edit. Explanations show how the query was interpreted. Results stream into a virtualized table; save the view or export when ready.
RBAC Matrix
- Right access, right place: Roles × permissions with targeted field overrides.
- Explainable controls: Disabled-with-reason tooltips reduce confusion.
Pricing Matrix
- Bulk without risk: Validation chips block unsafe applies.
- Audit-ready: Diff and receipts for each release window.
Pricing
- Scenario planning: See partner/territory impact before committing.
- History that helps: Trace when and why a price moved.
Strategy & principles: Make the rules visible and reversible
Key decisions & trade-offs
What we chose, what we gave up, and why it paid off.
-
Build metadata in the product (not a vendor queue)
DecisionIn-product schema editor with diff, lint, approval, rollback.
Trade-offWe own governance and tooling previously mediated by the vendor.
PayoffChanges ship at product speed; policy lives where work happens; history is automatic.
-
Insights that explain themselves
DecisionDeterministic rules + interpretable signals; ranked suggestions with citations; previewable diff.
Trade-offMore upfront modeling and UX for “why this change.”
PayoffHigher trust, higher acceptance, fewer escalations.
-
AI Search as typed, editable chips
DecisionParse natural language to typed filters with “why” popovers and safe fallbacks.
Trade-offMaintain a parsing layer, taxonomy, and partial-parse UX.
PayoffNovices form complex queries; experts get deterministic, shareable views.
-
One object model across modules
DecisionUnify IDs, history, and policy surfaces for titles, rights, prices, partners.
Trade-offHigher migration effort and naming discipline.
PayoffLess tool-hopping; fewer ambiguous states; simpler integrations.
-
Block only when unsafe
DecisionWarnings with “fix now” for non-critical issues; hard blocks for compliance risks.
Trade-offMore coaching UI to prevent alert fatigue.
PayoffMomentum without surprises; learning in context.
-
Receipts by default
DecisionEvery change writes who/when/why and before/after.
Trade-offEvent hygiene and small storage overhead.
PayoffAudits and handoffs run on evidence; disputes resolve quickly.
-
Bulk pricing with guardrails
DecisionValidation chips and publish summary before commit.
Trade-offExtra UI to visualize impact up front.
PayoffScale safely; cut downstream damage and rework.
-
Field-level RBAC where it matters
DecisionFine-grained controls for sensitive attributes (e.g., territorial pricing).
Trade-offMore admin surface and testing matrix.
PayoffIndependent team velocity with enforced boundaries.
-
Feed Console with guided retry
DecisionBucket errors by cause at the source record; retry same payload.
Trade-offInstrumentation and error taxonomy investment.
PayoffFaster recovery; fewer blind spots across systems.
-
Stream + virtualize results
DecisionStream rows and virtualize rendering; paginate only at scale thresholds.
Trade-offThreshold tuning and clear “still loading” states.
PayoffLarge catalogs remain responsive; search feels instant.
-
Preview → Apply → Undo everywhere
DecisionOne choreography across modules to reduce cognitive load.
Trade-offUX rigor to keep patterns consistent.
PayoffFaster onboarding; predictable operations.
-
Progressive disclosure, not mega-forms
DecisionShort, contextual steps with inline rules and readiness indicators.
Trade-offMore navigation states to design and test.
PayoffHigher completion rates; fewer user errors.
Design features mapped to screens
Business Partner
Guided setup for titles and contributors; required fields and policy checks inline; readiness status updates as you work.
Production
Milestones, assets, and dates on the same object model; late items surface as risks with suggested actions.
Pricing
Plan windows, simulate partner/territory effects, preview diffs, publish with a receipt.
Data feeds
Inbound/outbound logs tied to the source record; errors bucketed by cause; guided retry of the same payload.
Measured outcomes: Trust and speed, with receipts
Weeks → same-day schema changes; onboarding 2 weeks → 2.5 days; feed failures −43%.
−43%
Outbound feed failures
2.5d
Time to first independent title setup
+23pt
SUS (56 → 79)
−58%
“How do I publish/update price?” tickets
User reaction & adoption
- 74% completed a Title Setup in week one; 61% used AI Search; 38% saved a query.
- 68%of Insight proposals accepted after preview; most cited reasons: “shows me why,” “easy to undo.”
“I can see exactly what breaks and why, and the fix is right there.” — Editorial Lead
“Schema changes are a diff and an approval, not a sprint derail.” — Eng Manager
Not just screens; the operating model
Unified model
Titles, Contributors, Rights, Territories, Prices, Partners—one vocabulary across modules.
Governance
120+ validations; blocking only where unsafe; everything else warns with explainers.
Explainability
Propose → Preview → Apply → Undo with rule citations; immutable receipts.
Perf & scale
Chat → chips → indexed filters; virtualized grid; stream results and export.
Change management
Schema Builder replaces vendor queues; approvals and rollback lanes.
Measurement
Defined KPIs, instrumentation, ops dashboards; reviews moved from opinion to evidence.
What work looks like now
- Editorial Lead: System flags missing EU rights for an EU price; Insights previews the fix; apply; readiness turns green.
- Pricing Manager: Schedules a seasonal window; sees partner impact and history; applies with receipt.
- Ops Analyst: “hardcover AND price expiring in May AND EU rights pending” → chips compose filters → export table.
- Data Admin: Adds “Illustration Style” field → diff → approval → merged same day.
What surprised us
- People move faster when they can see consequences before they commit.
- Defaults reduce friction more than long configuration screens.
- Receipts improve reviews less debate, more agreement.
- AI skepticism is rational, transparency and undo turned skeptics into daily users.
Next & reflection
Next
- Deeper partner-level controls where policy differs by territory.
- More granular telemetry on proposal acceptance to refine ranking.
- Faster cross-module navigation (jump from a failed feed to the exact field).
Reflection
Designing for multiple roles on one model isn’t about hiding complexity; it’s about making it inspectable. By standardizing Preview → Apply → Undo and anchoring every action with a diff and a receipt, the system turned change from a risk into a routine. That’s what unlocked speed.