Esperia.
Introduction
Redesigned an enterprise data product with AI assistance, a unified platform that cut publishing time by 97%, enabled self serve 3x AI adoption, reduced support tickets by 68%, and improved documentation coverage from 6% to 71%.
Shipped
Lead Product Designer
Web - Enterprise SaaS
Web - Enterprise SaaS
Security Platform
4 months
Security Platform
4 months
Web - Enterprise SaaS
Project Overview
Esperia had spent four years building serious AI infrastructure, but data teams still operated across disconnected tools with no coherent layer tying the capabilities together. They lacked a unified workflow to create, configure, serve, and govern AI powered data products leaving fragmented tooling, poor discoverability, and zero visibility into how data was consumed downstream.
The Approach
I designed the end to end UX for the AI Product Studio across 6 modules, built in a new component for the design system and AI interaction and embedded AI into the platform from sprint one, to a 6 step creation wizard, a persistent AI Expert chat, configurable Knowledge Agents, and interactive data lineage.
The outcome
All 6 product modules shipped on Esperia Platform ONE. AI chat adoption rose 3×, creation errors dropped 64%, and platform NPS climbed from +8 to +44 within six months.
6/6
Modules Shipped
↑3×
AI Chat Adoption
+44
Platform NPS
The Approach
I designed the end to end UX for the AI Product Studio across 6 modules, built in a new component for the design system and AI interaction and embedded AI into the platform from sprint one, to a 6 step creation wizard, a persistent AI Expert chat, configurable Knowledge Agents, and interactive data lineage.
The outcome
All 6 product modules shipped on Esperia Platform ONE. AI chat adoption rose 3×, creation errors dropped 64%, and platform NPS climbed from +8 to +44 within six months.
6/6
Modules Shipped
↑3×
AI Chat Adoption
+44
Platform NPS
existing Research
Problems - Fill in the Gaps
‘’ I spend more time figuring out where the data lives than actually analysing it. It feels like archaeology, not analytics.’’
— VP Operations, Logistics enterprise client
01
Lack of common creation workflow
Building a data product meant jumping across separate tools and 14 steps for dataset registration, dashboards, ML models, and API config. Each step was disconnected,
02
AI was invisible to consumers
Non-technical stakeholders could view static dashboards but had no way to ask questions or get AI generated insights.
03
Zero data lineage visibility
There was no way to trace a product back to its source datasets and transformations, engineers had no view into downstream usage.
04
AI config was developer-only
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Design Process
Every design decision traced back to a user interview, a usability test, or a platform metric. Four phases, Discovery to GA launch, shipped in 4 months.
01
PHASE 01 · DISCOVERY
User Research & Problem Framing
18 interviews with data producers and consumers, plus a legacy tooling audit, to map where trust broke down.
User Interviews
Heuristic Audit
Journey Mapping
Competitive Analysis
Support Tickets
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
02
PHASE 02 · DEFINITION
Information architecture & Navigation
Card sorting and tree testing (81% success) validated a 5-space IA so nothing was buried.
IA Design
Card Sorting
Tree Testing
User Flow Mapping
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
High Fidelity Designs
Usability Testing
Iteration
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
01
PHASE 01 · DISCOVERY
User Research & Problem Framing
18 interviews with data producers and consumers, plus a legacy tooling audit, to map where trust broke down.
User Interviews
Heuristic Audit
Journey Mapping
Competitive Analysis
Support Tickets
02
PHASE 02 · DEFINITION
Information architecture & Navigation
Card sorting and tree testing (81% success) validated a 5-space IA so nothing was buried.
IA Design
Card Sorting
Tree Testing
User Flow Mapping
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
dISCOVERY
Research
Before designing, I synthesised existing signals from platform analytics, previous discovery sessions, and competitor analysis. This grounded design decisions rather than starting from intuition. Also used tools lIke JIRA, Whyser, Monday.com
18%
Data products with API serving
14 / 4 tools
Steps to publish (Legacy)
11%
Consumers attempting self-serve
6%
Products with lineage docs
+280% QoQ
“Can’t find dataset” tickets
+8
NPS (pre-redesign)
Understanding the User Group
18 user interviews across 3 personas - Data Producer, Data Consumer, Platform Admin at 6 enterprise clients spanning financial services, retail, logistics, and healthcare
THE BUILDER
Date Producer
Data analyst or engineer. Creates datasets, configures pipelines, trains models, builds dashboards. Needs a structured end to end creation flow that doesn't require jumping tools. Blocked by governance complexity, API setup.
The Stakeholder
Data Consumer
Business executive, product manager, operations lead. Views dashboards, wants to ask ad-hoc questions without waiting for analyst time. Needs: AI chat on dashboards, data freshness indicators, auto-generated reports.
The Operator
Platform Admin
Controls access, monitors consumption, manages governance policies. Needs lineage visibility, usage statistics (API tokens, interactions, consumer counts), and a clear data catalog with consistent metadata.
Understanding the User Group
18 user interviews across 3 personas - Data Producer, Data Consumer, Platform Admin at 6 enterprise clients spanning financial services, retail, logistics, and healthcare
THE BUILDER
Date Producer
Data analyst or engineer. Creates datasets, configures pipelines, trains models, builds dashboards. Needs a structured end to end creation flow that doesn't require jumping tools. Blocked by governance complexity, API setup.
The Stakeholder
Data Consumer
Business executive, product manager, operations lead. Views dashboards, wants to ask ad-hoc questions without waiting for analyst time. Needs: AI chat on dashboards, data freshness indicators, auto-generated reports.
The Operator
Platform Admin
Controls access, monitors consumption, manages governance policies. Needs lineage visibility, usage statistics (API tokens, interactions, consumer counts), and a clear data catalog with consistent metadata.
Heuristic Evaluation Findings
Evaluated the existing platform against Nielsen's 10 usability heuristics. Identified 47 issues across critical, high, medium, and low severity.
Critical
8 issues
Visibility of System Status
No progress indicators during data product creation. Users don't know if process is working or stuck.
Solution: Added step-by-step wizard with clear progress indicators and estimated time remaining for each phase.
Critical
6 issues
Error Prevention
No validation until final submission. Users make errors that only appear after 30+ minutes of configuration.
Solution: Implemented real-time validation with AI-powered error detection and prevention at each step.
High
12 issues
Help & Documentation
Documentation is technical and hard to find. No contextual help for complex fields.
Solution: Integrated AI Expert assistant providing contextual help, tooltips, and natural language guidance at every step.
High
9 issues
Recognition vs Recall
Users must remember complex configuration rules and field dependencies across multiple screens.
Solution: Smart field suggestions, auto-population based on context, and visual relationship indicators between fields.
Solutions
Decisions on the Screen
Each screen represents a core design decision, not just a layout choice, but a resolution to a specific user frustration or business goal.
6 Steps, Split panel with contextual AI attachment
A split view interface combined the dashboard canvas with a contextual panel for widgets, trained models and AI chat, surfacing key model details without leaving the workflow. Reduced data product creation time by 60%,
Knowledge Agent config with Live preview
Complex LLM settings were translated into guided controls with real-time preview, making AI configuration accessible to non technical users. API toggle and Public Access toggle give producers instant control over serving scope.
Column-Level API Serving
Enabled self-service API publishing for the first time, reducing engineering dependency while meeting enterprise security requirements through column-level access controls.
Product Detail and Lineage Nodes
A centralised product dashboard provided visibility into ownership, outputs, adoption, and usage metrics for the first time. API consumption metrics, helping producers understand how their data products were being used.
DESIGN SYSTEM
Addition of Lineage Node
Each screen represents a core design decision, not just a layout choice, but a resolution to a specific user frustration or business goal.
Interactive lineage graph
I along with the PLM have introduced new node-graph component built for the system, zoomable, schema-expandable nodes that trace a product back through its transformations and source datasets, with edge-level upstream/downstream highlighting.
impact
Changes after Launch
Measured at 6 months post-launch on Esperia Platform Unit, across pilot enterprise clients
6% → 71%
Lineage coverage
18% → 83%
API serving rate
14 → 6
Steps to publish
11% → 58%
Consumer self-serve
−60%
Dev handoff time
4.4 / 5
CSAT score
Take Away
Honest Retrospective
uNEXPECTED FINDINGS
Showing AI uncertainty increased trust
Users who saw the persistent disclaimer made better calibrated decisions. “Calibrated trust” became a formal UX concept in our AI principles.
uNEXPECTED FINDINGS
Producers wanted impact metrics more than any AI feature
The most mentioned feature was the Statistics panel proof that people were using what they built was more motivating than any AI capability.
Future
Test AI parameters with real users from sprint 1
There was no way to trace a product back to its source datasets and transformations, engineers had no view into downstream usage.
Future
Design the file architecture before the product
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Future
Design the file architecture before the product
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Earn engineering trust by designing for engineers
Dev ready files, status systems, and pattern docs made design a partner to engineering, not a bottleneck.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Earn engineering trust by designing for engineers
Dev ready files, status systems, and pattern docs made design a partner to engineering, not a bottleneck.
Esperia.
Introduction
Redesigned an enterprise data product with AI assistance, a unified platform that cut publishing time by 97%, enabled self serve 3x AI adoption, reduced support tickets by 68%, and improved documentation coverage from 6% to 71%.
Shipped
Lead Product Designer
Web - Enterprise SaaS
Web - Enterprise SaaS
Security Platform
4 months
Security Platform
4 months
Web - Enterprise SaaS
Project Overview
Esperia had spent four years building serious AI infrastructure, but data teams still operated across disconnected tools with no coherent layer tying the capabilities together. They lacked a unified workflow to create, configure, serve, and govern AI powered data products leaving fragmented tooling, poor discoverability, and zero visibility into how data was consumed downstream.
The Approach
I designed the end to end UX for the AI Product Studio across 6 modules, built in a new component for the design system and AI interaction and embedded AI into the platform from sprint one, to a 6 step creation wizard, a persistent AI Expert chat, configurable Knowledge Agents, and interactive data lineage.
The outcome
All 6 product modules shipped on Esperia Platform ONE. AI chat adoption rose 3×, creation errors dropped 64%, and platform NPS climbed from +8 to +44 within six months.
6/6
Modules Shipped
↑3×
AI Chat Adoption
+44
Platform NPS
The Approach
I designed the end to end UX for the AI Product Studio across 6 modules, built in a new component for the design system and AI interaction and embedded AI into the platform from sprint one, to a 6 step creation wizard, a persistent AI Expert chat, configurable Knowledge Agents, and interactive data lineage.
The outcome
All 6 product modules shipped on Esperia Platform ONE. AI chat adoption rose 3×, creation errors dropped 64%, and platform NPS climbed from +8 to +44 within six months.
6/6
Modules Shipped
↑3×
AI Chat Adoption
+44
Platform NPS
existing Research
Problems -
Fill in the Gaps
‘’ I spend more time figuring out where the data lives than actually analysing it. It feels like archaeology, not analytics.’’
— VP Operations, Logistics enterprise client
01
Lack of common creation workflow
Building a data product meant jumping across separate tools and 14 steps for dataset registration, dashboards, ML models, and API config. Each step was disconnected,
02
AI was invisible to consumers
Non-technical stakeholders could view static dashboards but had no way to ask questions or get AI generated insights.
03
Zero data lineage visibility
There was no way to trace a product back to its source datasets and transformations, engineers had no view into downstream usage.
04
AI config was developer-only
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Design Process
Every design decision traced back to a user interview, a usability test, or a platform metric. Four phases, Discovery to GA launch, shipped in 4 months.
01
PHASE 01 · DISCOVERY
User Research & Problem Framing
18 interviews with data producers and consumers, plus a legacy tooling audit, to map where trust broke down.
User Interviews
Heuristic Audit
Journey Mapping
Competitive Analysis
Support Tickets
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
02
PHASE 02 · DEFINITION
Information architecture & Navigation
Card sorting and tree testing (81% success) validated a 5-space IA so nothing was buried.
IA Design
Card Sorting
Tree Testing
User Flow Mapping
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
High Fidelity Designs
Usability Testing
Iteration
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
01
PHASE 01 · DISCOVERY
User Research & Problem Framing
18 interviews with data producers and consumers, plus a legacy tooling audit, to map where trust broke down.
User Interviews
Heuristic Audit
Journey Mapping
Competitive Analysis
Support Tickets
02
PHASE 02 · DEFINITION
Information architecture & Navigation
Card sorting and tree testing (81% success) validated a 5-space IA so nothing was buried.
IA Design
Card Sorting
Tree Testing
User Flow Mapping
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
dISCOVERY
Research
Before designing, I synthesised existing signals from platform analytics, previous discovery sessions, and competitor analysis. This grounded design decisions rather than starting from intuition. Also used tools lIke JIRA, Whyser, Monday.com
18%
Data products with API serving
14 / 4 tools
Steps to publish (Legacy)
11%
Consumers attempting self-serve
6%
Products with lineage docs
+280% QoQ
“Can’t find dataset” tickets
+8
NPS (pre-redesign)
Understanding the User Group
18 user interviews across 3 personas - Data Producer, Data Consumer, Platform Admin at 6 enterprise clients spanning financial services, retail, logistics, and healthcare
THE BUILDER
Date Producer
Data analyst or engineer. Creates datasets, configures pipelines, trains models, builds dashboards. Needs a structured end to end creation flow that doesn't require jumping tools. Blocked by governance complexity, API setup.
The Stakeholder
Data Consumer
Business executive, product manager, operations lead. Views dashboards, wants to ask ad-hoc questions without waiting for analyst time. Needs: AI chat on dashboards, data freshness indicators, auto-generated reports.
The Operator
Platform Admin
Controls access, monitors consumption, manages governance policies. Needs lineage visibility, usage statistics (API tokens, interactions, consumer counts), and a clear data catalog with consistent metadata.
Understanding the User Group
18 user interviews across 3 personas - Data Producer, Data Consumer, Platform Admin at 6 enterprise clients spanning financial services, retail, logistics, and healthcare
THE BUILDER
Date Producer
Data analyst or engineer. Creates datasets, configures pipelines, trains models, builds dashboards. Needs a structured end to end creation flow that doesn't require jumping tools. Blocked by governance complexity, API setup.
The Stakeholder
Data Consumer
Business executive, product manager, operations lead. Views dashboards, wants to ask ad-hoc questions without waiting for analyst time. Needs: AI chat on dashboards, data freshness indicators, auto-generated reports.
The Operator
Platform Admin
Controls access, monitors consumption, manages governance policies. Needs lineage visibility, usage statistics (API tokens, interactions, consumer counts), and a clear data catalog with consistent metadata.
Heuristic Evaluation Findings
Evaluated the existing platform against Nielsen's 10 usability heuristics. Identified 47 issues across critical, high, medium, and low severity.
Critical
8 issues
Visibility of System Status
No progress indicators during data product creation. Users don't know if process is working or stuck.
Solution: Added step-by-step wizard with clear progress indicators and estimated time remaining for each phase.
Critical
6 issues
Error Prevention
No validation until final submission. Users make errors that only appear after 30+ minutes of configuration.
Solution: Implemented real-time validation with AI-powered error detection and prevention at each step.
High
12 issues
Help & Documentation
Documentation is technical and hard to find. No contextual help for complex fields.
Solution: Integrated AI Expert assistant providing contextual help, tooltips, and natural language guidance at every step.
High
9 issues
Recognition vs Recall
Users must remember complex configuration rules and field dependencies across multiple screens.
Solution: Smart field suggestions, auto-population based on context, and visual relationship indicators between fields.
Solutions
Decisions on the Screen
Each screen represents a core design decision, not just a layout choice, but a resolution to a specific user frustration or business goal.
6 Steps, Split panel with contextual AI attachment
A split view interface combined the dashboard canvas with a contextual panel for widgets, trained models and AI chat, surfacing key model details without leaving the workflow. Reduced data product creation time by 60%,
Knowledge Agent config with Live preview
Complex LLM settings were translated into guided controls with real-time preview, making AI configuration accessible to non technical users. API toggle and Public Access toggle give producers instant control over serving scope.
Column-Level API Serving
Enabled self-service API publishing for the first time, reducing engineering dependency while meeting enterprise security requirements through column-level access controls.
Product Detail and Lineage Nodes
A centralised product dashboard provided visibility into ownership, outputs, adoption, and usage metrics for the first time. API consumption metrics, helping producers understand how their data products were being used.
DESIGN SYSTEM
Addition of Lineage Node
Each screen represents a core design decision, not just a layout choice, but a resolution to a specific user frustration or business goal.
Interactive lineage graph
I along with the PLM have introduced new node-graph component built for the system, zoomable, schema-expandable nodes that trace a product back through its transformations and source datasets, with edge-level upstream/downstream highlighting.
impact
Changes after Launch
Measured at 6 months post-launch on Esperia Platform Unit, across pilot enterprise clients
6% → 71%
Lineage coverage
18% → 83%
API serving rate
14 → 6
Steps to publish
11% → 58%
Consumer self-serve
−60%
Dev handoff time
4.4 / 5
CSAT score
Take Away
Honest
Retrospective
uNEXPECTED FINDINGS
Showing AI uncertainty increased trust
Users who saw the persistent disclaimer made better calibrated decisions. “Calibrated trust” became a formal UX concept in our AI principles.
uNEXPECTED FINDINGS
Producers wanted impact metrics more than any AI feature
The most mentioned feature was the Statistics panel proof that people were using what they built was more motivating than any AI capability.
Future
Test AI parameters with real users from sprint 1
There was no way to trace a product back to its source datasets and transformations, engineers had no view into downstream usage.
Future
Design the file architecture before the product
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Future
Design the file architecture before the product
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Earn engineering trust by designing for engineers
Dev ready files, status systems, and pattern docs made design a partner to engineering, not a bottleneck.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Earn engineering trust by designing for engineers
Dev ready files, status systems, and pattern docs made design a partner to engineering, not a bottleneck.
Esperia.
Introduction
Redesigned an enterprise data product with AI assistance, a unified platform that cut publishing time by 97%, enabled self serve 3x AI adoption, reduced support tickets by 68%, and improved documentation coverage from 6% to 71%.
Shipped
Lead Product Designer
Web - Enterprise SaaS
Web - Enterprise SaaS
Security Platform
4 months
Security Platform
4 months
Web - Enterprise SaaS
Project Overview
Esperia had spent four years building serious AI infrastructure, but data teams still operated across disconnected tools with no coherent layer tying the capabilities together. They lacked a unified workflow to create, configure, serve, and govern AI powered data products leaving fragmented tooling, poor discoverability, and zero visibility into how data was consumed downstream.
The Approach
I designed the end to end UX for the AI Product Studio across 6 modules, built in a new component for the design system and AI interaction and embedded AI into the platform from sprint one, to a 6 step creation wizard, a persistent AI Expert chat, configurable Knowledge Agents, and interactive data lineage.
The outcome
All 6 product modules shipped on Esperia Platform ONE. AI chat adoption rose 3×, creation errors dropped 64%, and platform NPS climbed from +8 to +44 within six months.
6/6
Modules Shipped
↑3×
AI Chat Adoption
+44
Platform NPS
The Approach
I designed the end to end UX for the AI Product Studio across 6 modules, built in a new component for the design system and AI interaction and embedded AI into the platform from sprint one, to a 6 step creation wizard, a persistent AI Expert chat, configurable Knowledge Agents, and interactive data lineage.
The outcome
All 6 product modules shipped on Esperia Platform ONE. AI chat adoption rose 3×, creation errors dropped 64%, and platform NPS climbed from +8 to +44 within six months.
6/6
Modules Shipped
↑3×
AI Chat Adoption
+44
Platform NPS
existing Research
Problems - Fill in the Gaps
‘’ I spend more time figuring out where the data lives than actually analysing it. It feels like archaeology, not analytics.’’
— VP Operations, Logistics enterprise client
01
Lack of common creation workflow
Building a data product meant jumping across separate tools and 14 steps for dataset registration, dashboards, ML models, and API config. Each step was disconnected.
02
AI was invisible to consumers
Non-technical stakeholders could view static dashboards but had no way to ask questions or get AI generated insights.
03
Zero data lineage visibility
There was no way to trace a product back to its source datasets and transformations, engineers had no view into downstream usage.
04
AI config was developer-only
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Design Process
Every design decision traced back to a user interview, a usability test, or a platform metric. Four phases, Discovery to GA launch, shipped in 4 months.
01
PHASE 01 · DISCOVERY
User Research & Problem Framing
18 interviews with data producers and consumers, plus a legacy tooling audit, to map where trust broke down.
User Interviews
Heuristic Audit
Journey Mapping
Competitive Analysis
Support Tickets
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
02
PHASE 02 · DEFINITION
Information architecture & Navigation
Card sorting and tree testing (81% success) validated a 5-space IA so nothing was buried.
IA Design
Card Sorting
Tree Testing
User Flow Mapping
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
High Fidelity Designs
Usability Testing
Iteration
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
01
PHASE 01 · DISCOVERY
User Research & Problem Framing
18 interviews with data producers and consumers, plus a legacy tooling audit, to map where trust broke down.
User Interviews
Heuristic Audit
Journey Mapping
Competitive Analysis
Support Tickets
02
PHASE 02 · DEFINITION
Information architecture & Navigation
Card sorting and tree testing (81% success) validated a 5-space IA so nothing was buried.
IA Design
Card Sorting
Tree Testing
User Flow Mapping
03
PHASE 03 · DESIGN & TEST
Design system, Iterative design + usability testing
Four rounds of moderated testing surfaced the AI-surface fires wizard length, persistent chat, exposed LLM controls.
High Fidelity Designs
Usability Testing
Iteration
04
PHASE 04 · SHIP & SCALE
Rollout across all modules
Three dev-review rounds and DEV-Ready docs. All 6 modules shipped, then two follow up iterations.
Dev Handoff
Accessibility
Key metrics evaluation
dISCOVERY
Research
Before designing, I synthesised existing signals from platform analytics, previous discovery sessions, and competitor analysis. This grounded design decisions rather than starting from intuition. Also used tools lIke JIRA, Whyser, Monday.com
18%
Data products with
API serving
14 / 4 tools
Steps to publish (Legacy)
11%
Consumers attempting
self-serve
6%
Products with
lineage docs
+280% QoQ
“Can’t find dataset” tickets
+8
NPS (pre-redesign)
Understanding the User Group
18 user interviews across 3 personas - Data Producer, Data Consumer, Platform Admin at 6 enterprise clients spanning financial services, retail, logistics, and healthcare
THE BUILDER
Date Producer
Data analyst or engineer. Creates datasets, configures pipelines, trains models, builds dashboards. Needs a structured end to end creation flow that doesn't require jumping tools. Blocked by governance complexity, API setup.
The Stakeholder
Data Consumer
Business executive, product manager, operations lead. Views dashboards, wants to ask ad-hoc questions without waiting for analyst time. Needs: AI chat on dashboards, data freshness indicators, auto-generated reports.
The Operator
Platform Admin
Controls access, monitors consumption, manages governance policies. Needs lineage visibility, usage statistics (API tokens, interactions, consumer counts), and a clear data catalog with consistent metadata.
Understanding the User Group
18 user interviews across 3 personas - Data Producer, Data Consumer, Platform Admin at 6 enterprise clients spanning financial services, retail, logistics, and healthcare
THE BUILDER
Date Producer
Data analyst or engineer. Creates datasets, configures pipelines, trains models, builds dashboards. Needs a structured end to end creation flow that doesn't require jumping tools. Blocked by governance complexity, API setup.
The Stakeholder
Data Consumer
Business executive, product manager, operations lead. Views dashboards, wants to ask ad-hoc questions without waiting for analyst time. Needs: AI chat on dashboards, data freshness indicators, auto-generated reports.
The Operator
Platform Admin
Controls access, monitors consumption, manages governance policies. Needs lineage visibility, usage statistics (API tokens, interactions, consumer counts), and a clear data catalog with consistent metadata.
Heuristic Evaluation Findings
Evaluated the existing platform against Nielsen's 10 usability heuristics. Identified 47 issues across critical, high, medium, and low severity.
Critical
8 issues
Visibility of System Status
No progress indicators during data product creation. Users don't know if process is working or stuck.
Solution: Added step-by-step wizard with clear progress indicators and estimated time remaining for each phase.
Critical
6 issues
Error Prevention
No validation until final submission. Users make errors that only appear after 30+ minutes of configuration.
Solution: Implemented real-time validation with AI-powered error detection and prevention at each step.
High
12 issues
Help & Documentation
Documentation is technical and hard to find. No contextual help for complex fields.
Solution: Integrated AI Expert assistant providing contextual help, tooltips, and natural language guidance at every step.
High
9 issues
Recognition vs Recall
Users must remember complex configuration rules and field dependencies across multiple screens.
Solution: Smart field suggestions, auto-population based on context, and visual relationship indicators between fields.
Solutions
Decisions on the Screen
Each screen represents a core design decision, not just a layout choice, but a resolution to a specific user frustration or business goal.
6 Steps, Split panel with contextual AI attachment
A split view interface combined the dashboard canvas with a contextual panel for widgets, trained models and AI chat, surfacing key model details without leaving the workflow. Reduced data product creation time by 60%,
Knowledge Agent config with Live preview
Complex LLM settings were translated into guided controls with real-time preview, making AI configuration accessible to non technical users. API toggle and Public Access toggle give producers instant control over serving scope.
Column-Level API Serving
Enabled self-service API publishing for the first time, reducing engineering dependency while meeting enterprise security requirements through column-level access controls.
Product Detail and Lineage Nodes
A centralised product dashboard provided visibility into ownership, outputs, adoption, and usage metrics for the first time. API consumption metrics, helping producers understand how their data products were being used.
DESIGN SYSTEM
Addition of Lineage Node
Each screen represents a core design decision, not just a layout choice, but a resolution to a specific user frustration or business goal.
Interactive lineage graph
I along with the PLM have introduced new node-graph component built for the system, zoomable, schema-expandable nodes that trace a product back through its transformations and source datasets, with edge-level upstream/downstream highlighting.
impact
Changes after Launch
Measured at 6 months post-launch on Esperia Platform Unit, across pilot enterprise clients
6% → 71%
Lineage coverage
18% → 83%
API serving rate
14 → 6
Steps to publish
11% → 58%
Consumer self-serve
−60%
Dev handoff time
4.4 / 5
CSAT score
Take Away
Honest Retrospective
uNEXPECTED FINDINGS
Showing AI uncertainty increased trust
Users who saw the persistent disclaimer made better calibrated decisions. “Calibrated trust” became a formal UX concept in our AI principles.
uNEXPECTED FINDINGS
Showing AI uncertainty increased trust
The most mentioned feature was the Statistics panel proof that people were using what they built was more motivating than any AI capability.
Future
Test AI parameters with real users from sprint 1
There was no way to trace a product back to its source datasets and transformations, engineers had no view into downstream usage.
Future
Design the file architecture before the product
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Future
Design the file architecture before the product
Setting up a Knowledge Agent or Trained Model required engineering. Non-technical producers couldn't tune LLM parameters independently.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Earn engineering trust by designing for engineers
Dev ready files, status systems, and pattern docs made design a partner to engineering, not a bottleneck.
Leadership
Constraints documentation is a leadership act
Getting PM, engineering, and legal to sign off on the constraint register in week 1 prevented the most design rework, a shared boundary, not a debate.
Leadership
Earn engineering trust by designing for engineers
Dev ready files, status systems, and pattern docs made design a partner to engineering, not a bottleneck.