72% of users caught model issues faster with a simpler way to see AI health

72% of users caught model issues faster with a simpler way to see AI health

Services

Product design · UX research · Value proposition design · Interaction design · Cross-functional collaboration (engineering & data science)

Description

Many organisations struggle to maintain ML models after deployment, errors accumulate, performance drifts, and teams lose visibility into whether models can still be trusted. Coda was designed to make model health understandable and actionable through a simple, accessible interface for monitoring AI systems.

Many organisations struggle to maintain ML models after deployment, errors accumulate, performance drifts, and teams lose visibility into whether models can still be trusted. Coda was designed to make model health understandable and actionable through a simple, accessible interface for monitoring AI systems.

Impact

  • 72% of users felt it helped them detect challenges faster

  • ~60% increase in usage

  • Took the product from discovery to production, including refining the value proposition

  • Simplified complex ML concepts into intuitive, visual workflows

Year

2025

Timeline

Growth

Location

London, UK

Left, screenshots of some rounds of interviews captured in Dovetail. Right, a screenshot of Coda when the work was started, it was not built for purpose.

Problem space

The problems this work had to make visible

The problems this work had to make visible

The problems this work had to make visible

Coda had to solve for both a technical gap in model maintenance and a research gap in understanding the people relying on it.

01

Maintaining models

Once external consultants left, organisations struggled to maintain models, errors crept in and performance grew inconsistent. Operational teams simply lacked tools built for ongoing maintenance.

Once external consultants left, organisations struggled to maintain models, errors crept in and performance grew inconsistent. Operational teams simply lacked tools built for ongoing maintenance.

Once external consultants left, organisations struggled to maintain models, errors crept in and performance grew inconsistent. Operational teams simply lacked tools built for ongoing maintenance.

02

Team Dynamics and Understanding the User

An early version of Coda existed, but the team didn't yet understand its primary users, LiveOps engineers, well enough to serve them. This was the first time a designer joined to shape the product around real user needs.

An early version of Coda existed, but the team didn't yet understand its primary users, LiveOps engineers, well enough to serve them. This was the first time a designer joined to shape the product around real user needs.

An early version of Coda existed, but the team didn't yet understand its primary users, LiveOps engineers, well enough to serve them. This was the first time a designer joined to shape the product around real user needs.

03

Not Built for Purpose and Lacking Trust

Engineers needed to triage errors by severity fast, and wanted flexible timeframes to match different model retraining schedules. They also wanted visibility into business impact, not just technical metrics, and clear explanations to build trust in a new system.

Engineers needed to triage errors by severity fast, and wanted flexible timeframes to match different model retraining schedules. They also wanted visibility into business impact, not just technical metrics, and clear explanations to build trust in a new system.

Engineers needed to triage errors by severity fast, and wanted flexible timeframes to match different model retraining schedules. They also wanted visibility into business impact, not just technical metrics, and clear explanations to build trust in a new system.

Concept exploration explored when ideating about ways we could showcase the data,

Contribution

Where I created leverage across the work

Where I created leverage across the work

As design lead, I embedded research into the team's rhythm and turned what we learned into a dashboard engineers could trust.

01

Research & Research Culture

Ran regular qualitative interviews with users and stakeholders, testing new features against real needs rather than assumptions. Brought the whole team into research playbacks, building a genuinely research-driven culture.

Ran regular qualitative interviews with users and stakeholders, testing new features against real needs rather than assumptions. Brought the whole team into research playbacks, building a genuinely research-driven culture.

Ran regular qualitative interviews with users and stakeholders, testing new features against real needs rather than assumptions. Brought the whole team into research playbacks, building a genuinely research-driven culture.

02

Experience design

Designed a hierarchical dashboard that surfaces critical errors first, colour-coded for an instant read on pipeline health. Trend views gave engineers a fast, familiar way to judge run health over time.

Designed a hierarchical dashboard that surfaces critical errors first, colour-coded for an instant read on pipeline health. Trend views gave engineers a fast, familiar way to judge run health over time.

Designed a hierarchical dashboard that surfaces critical errors first, colour-coded for an instant read on pipeline health. Trend views gave engineers a fast, familiar way to judge run health over time.

03

Flexible, Business-Aware Monitoring

Flexible time-period views kept Coda relevant across clients retraining models on different schedules. Surfaced business outcomes like model adherence alongside technical metrics, directly closing the trust gap research uncovered.

Flexible time-period views kept Coda relevant across clients retraining models on different schedules. Surfaced business outcomes like model adherence alongside technical metrics, directly closing the trust gap research uncovered.

Flexible time-period views kept Coda relevant across clients retraining models on different schedules. Surfaced business outcomes like model adherence alongside technical metrics, directly closing the trust gap research uncovered.

© All right reserved

© All right reserved