WindSure: Designing an Predictive Maintenance MVP

RadiAI : Designing an AI-Powered Chest X-Ray Analysis Application

To build an AI-powered platform that combines sensor data and drone inspection results into a single dashboard. This gives operators a clear picture of turbine health before issues develop.

Background

Wind turbines face extreme conditions, and each outage can cost operators up to $80,000 per day. Outages disrupt power supply, cause lost revenue, and require costly repairs.


Operators use predictive maintenance and drone inspections to keep track of turbine health. But there isn't one tool that combines both methods to predict failures and spot external damage together.


We set out to create an AI-powered platform(MVP) that gives both executives and engineers a reliable, easy-to-use overview of turbine health.


This project served as a proof-of-concept for our ability to deploy high-performance Nvidia GPUs, utilising state-of-the-art AI models to ingest and analyse high-volume sensor and visual data efficiently.

This project was recognised with the Star Team Award at Genpact for its intuitive UX, speed, precision, and innovation delivered in just 5 weeks.

Design Approach & Tight Timelines

This was a high-stakes project from day one. The CTO was directly involved, and we had just five weeks to deliver a working demo. My job was to shape the design so it looked good and clearly reflected real work processes.


To keep up the pace, we set three main milestones. First, we spent three days on research and workshops to find out what was needed. Next, we used two days to test wireframes and get feedback. Finally, we quickly built high-fidelity prototypes to guide the development.


We aimed to have a full low-fidelity wireframe and a couple of high-fidelity screens ready in the first week. This way, development could start while we continued designing the rest of the screens alongside.

This was a high-stakes project from day one. The CTO was directly involved, and we had just five weeks to deliver a working demo. My job was to shape the design so it looked good and clearly reflected real work processes.


To keep up the pace, we set three main milestones. First, we spent three days on research and workshops to find out what was needed. Next, we used two days to test wireframes and get feedback. Finally, we quickly built high-fidelity prototypes to guide the development.


We aimed to have a full low-fidelity wireframe and a couple of high-fidelity screens ready in the first week. This way, development could start while we continued designing the rest of the screens alongside.

This was a high-stakes project from day one. The CTO was directly involved, and we had just five weeks to deliver a working demo. My job was to shape the design so it looked good and clearly reflected real work processes.


To keep up the pace, we set three main milestones. First, we spent three days on research and workshops to find out what was needed. Next, we used two days to test wireframes and get feedback. Finally, we quickly built high-fidelity prototypes to guide the development.


We aimed to have a full low-fidelity wireframe and a couple of high-fidelity screens ready in the first week. This way, development could start while we continued designing the rest of the screens alongside.

Research

I started out knowing little about turbine maintenance, but was able to quickly bridge the gap. The project lead and I spoke with site engineers from halfway across the world, who showed us their daily routines and shared their insights. We gathered a lot of information, but with time short, we had to focus on what was most important. Our aim was to design a demo that felt real and practical, showing off our strengths in AI and machine learning without getting lost in the details.


We also looked back at our previous computer vision projects to find useful interaction patterns that could help guide this work.

I started out knowing little about turbine maintenance, but was able to quickly bridge the gap. The project lead and I spoke with site engineers from halfway across the world, who showed us their daily routines and shared their insights. We gathered a lot of information, but with time short, we had to focus on what was most important. Our aim was to design a demo that felt real and practical, showing off our strengths in AI and machine learning without getting lost in the details.


We also looked back at our previous computer vision projects to find useful interaction patterns that could help guide this work.

I started out knowing little about turbine maintenance, but was able to quickly bridge the gap. The project lead and I spoke with site engineers from halfway across the world, who showed us their daily routines and shared their insights. We gathered a lot of information, but with time short, we had to focus on what was most important. Our aim was to design a demo that felt real and practical, showing off our strengths in AI and machine learning without getting lost in the details.


We also looked back at our previous computer vision projects to find useful interaction patterns that could help guide this work.

Some of the key insights that informed the Solution

Different Roles = Different Definitions of “Value”

Insight: Executives focus on financial impact and risk, while engineers prioritize real-time status and maintenance details. Both value fleet performance, but their definitions of value are distinct.
Impact on Design: We created role-based dashboards with tailored KPIs and entry points to support effective decision-making. Executives saw ROI, downtime cost, and risk forecasts upfront, while engineers accessed alarms, task queues, and sensor trends all within a shared navigation structure.

Insight: Executives focus on financial impact and risk, while engineers prioritize real-time status and maintenance details. Both value fleet performance, but their definitions of value are distinct.
Impact on Design: We created role-based dashboards with tailored KPIs and entry points to support effective decision-making. Executives saw ROI, downtime cost, and risk forecasts upfront, while engineers accessed alarms, task queues, and sensor trends all within a shared navigation structure.

Insight: Executives focus on financial impact and risk, while engineers prioritize real-time status and maintenance details. Both value fleet performance, but their definitions of value are distinct.
Impact on Design: We created role-based dashboards with tailored KPIs and entry points to support effective decision-making. Executives saw ROI, downtime cost, and risk forecasts upfront, while engineers accessed alarms, task queues, and sensor trends all within a shared navigation structure.

Big Picture + Drill-Down Improves Decision Making

Insight: Executives require a fleet-wide overview, whereas engineers need detailed, turbine-level insights to manage daily operations.
For the design, we organized information from the fleet level down to each site and turbine. This way, users could start with the big picture and then dig into the details as needed.

“Tech Output” ≠ “Actionable Insight”

Insight: Experts trust sensor data, but leadership struggles to link it to business outcomes. This gap makes justifying predictive maintenance challenging.


We turned technical results into business-focused visuals, like showing a predicted gearbox failure as a potential $1.2 million loss. Each prediction included a confidence level, helping connect engineering details to business decisions.

Building Empathy

Based on the interview insights, I quickly developed a Proto Persona to build empathy for the primary user group, i.e. Operations Engineer and Head of Operations and Maintenance. This process was accelerated by using GenAI tools for rapid research synthesis and persona creation

Translating Research into Structure

The most challenging part was synthesising our research into a compelling data narrative. The design had to meet two critical requirements: establishing our expertise in AI while ensuring the solution remained accessible and relatable to our audience.


To harmonise these constraints, we used a persona-centric approach, framing content around the user's core need: 'What is the purpose of predictive maintenance for me?' This allowed us to filter and determine the exact KPIs necessary for the MVP.


We then decided on progressive information disclosure. We started with the big picture (fleet-level insights) before allowing users to dive into individual turbines. This storytelling approach first provides an overview of how our AI creates preemptive alerts, demonstrating value. Users can then dive deeper into the turbine and alerts view to better understand the problem, troubleshoot, and schedule maintenance.


Below are the finalised sitemap showing data points and content organised for each screen:

Wireframe

Lorem Ipsum

Dashboard

We designed persona-centric dashboards that ensured relevance for every user. Executives viewed high-level financial impact metrics, while the Operations Engineer had immediate access to detailed maintenance and performance KPIs.

Turbine Monitoring

The Turbine Monitoring screen enables proactive maintenance. Here operators can immediately access predictive and real-time alerts generated by the AI, quickly review any detected damages, and monitor detailed performance metrics to prevent issues before they cause downtime.

Detected Damage Details

The Damage Detection Screen allows the Operations Engineer to move from insight to action. Here, the engineer can access detailed views of detected damage and verify the AI's analysis. If the AI's detection is confirmed, the engineer can seamlessly schedule the necessary repair directly from this interface

Drone Inspection

The Drone Inspection Screen serves as the hub for visual asset assessment. It allows the engineer to review raw drone footage alongside the AI's visual analysis, clearly highlighting any detected anomalies or damages. From this centralised view, the engineer can seamlessly raise official repair requests for confirmed issues.

Reflection

After gathering insights, I realised the solution needed to fit naturally into how radiologists already work. To do that, I focused on three design