The problem
Two maintainers could hit the same fault and land on different fixes.
When a fault code came up, a maintainer had a paper manual, their own memory and notes from previous repairs. What they often lacked was the aircraft's own repair history. It lived in logs that were often paper and did not reliably travel with the aircraft, so there was no way to know in advance whether it would be there.
That made the fix depend on who was doing the work. An experienced maintainer could draw on residual knowledge from years of repairs. Someone new to the industry could not.
My role and the team
I led two engineers and an AI vendor as technical product owner and designed every screen.
Honeywell had already identified the opportunity and lined up an AI/ML vendor, and the program was brought to me to lead. I led two engineers, managed the relationship with the vendor and designed the experience end to end. The vendor built the model that ingested the manual and repair logs. My engineers built the app to my designs and the backend.
This was my first time doing a technical product owner's job. I was still a researcher by title and nobody on the project reported to me, so I directed the work without managing the people.
The constraints were simple to state and hard to design around: about six months, two engineers, a cross-platform build for iPhone and Android and an outside vendor whose model was good at surfacing candidate fixes but not at certainty.
Designing for trust
I grouped suggestions by source and confidence so a maintainer could see where a recommendation came from and how sure the system was.
The easy version of this product would have shown a maintainer "the answer." A maintainer's judgment is the safety mechanism, so the screen had to show where a recommendation came from and how sure the system was.
I designed the flow around two sources, the manufacturer's manual and the aircraft's repair logs, shown in three groups on the Observation Details screen. I worked out this structure together with the vendor.
-
Prescribed by Manual
Suggestions prescribed directly by the manual
-
Suggested by MX History
Suggestions drawn from repair history, with a confidence score
-
Manual and MX History
Suggestions where the two agreed, with a confidence score
Selecting a suggestion opens the Troubleshooting Steps screen, with fault confirmation and fault isolation steps from the manual, expandable and collapsible. The expandable layout followed the maintainer sessions described below. Sessions are organized by aircraft, and each session lists its fault observations.
Explainability is not a feature to add once the recommendation works. It is the first thing to design.
Working with the vendor
The vendor's model produced unreadable text from the repair manuals, so I pushed back until they reprocessed them.
The app could only be as trustworthy as the suggestions it showed, so my engineers and I reviewed the model's output with the vendor. The biggest problem was the manufacturers' repair manuals. The model had trouble reading them, and the text it produced was often unclear or full of random characters. We sent it back each time, and the vendor reprocessed the manuals. The text got cleaner with every round.
What maintainers told us
Four maintainers checked the app's suggestions against their own approach before acting.
I came to this project knowing nothing about maintainers, so I wanted to find out whether they would trust the recommendations before the program invested further. Before we demonstrated it to airlines, I tested the working app one-on-one with four maintainers, all Honeywell customers. The sessions ran remotely through UserTesting and were recorded. Each maintainer used the app to evaluate a made-up aircraft issue that I gave them.
The response was consistent. Maintainers looked at what the app suggested and checked it against how they would normally approach the repair before acting. Nobody was going to hand a fault over to a recommendation engine sight unseen. That supported the source-and-confidence design. My read of the sessions is that they understood what we were trying to do and could see the value; the main issue was trusting the AI agent.
Maintainers also found the manual text hard to read with all of it on one page. I had expected that and needed a little data to back it up. The sessions supplied it, and I worked with the team to break the manual content into pieces that expand and collapse on demand.
What we built
A working iPhone and Android app that showed its sources and its confidence.
The walkthrough below is a screen recording of the live app with dummy data. It shows the troubleshooting flow end to end: a maintainer opening a session, reviewing suggestions grouped by source and confidence and working through the recommended steps.
What happened
I demonstrated the working app to two airlines in South Korea. Both passed, and the program was shelved.
In May 2018, I traveled to South Korea and presented the working iPhone and Android app to two airlines, Asiana and Korean Air, as a possible solution. It was a live demonstration. Both airlines passed. The program was later shelved, and I moved on to lead a different initiative, engine data access.
What I learned
Design explainability first, expect expert users to check the tool and remember that being early is not being wrong.
-
Explainability comes first
With users whose judgment is the actual safety mechanism, it has to be the thing you design first.
-
Trust is earned in the interface
The maintainers were not resisting AI. They were doing what a good maintainer should do with any new tool, which is refuse to trust it blindly.
-
Being early is not the same as being wrong
In 2018 the industry was not yet comfortable with confidence-scored AI recommendations. They are closer to standard practice now.
-
What I would do differently
I would have run another round of testing with maintainers from the engine data project. The team was targeting large airlines, and business aviation may have been a better fit, though I never tested that.
