Belkin Phyn
AI Water iot
The Situation
The team: 1 product manager, ~30 engineers, 3 data scientists
Belkin was building IoT devices that used machine learning engines to pair flow/current with resonance to predict which devices were in use and detect anomalies. They knew from experience that users often have a hard time setting things up. I was the sole designer creating the installation routine and app for their water device.
My role: Lead designer
Timeline: 1.5 years
Learning ML
I had helped put together Stanford's Precourt Energy Institute Energy and Behavior Change conference in 2009, so I was familiar with early research into using electrical wire resonance and voltage drop to predict device use. By 2012, using the same concept, Belkin's data scientists had an algorithm that could measure flow and resonance to predict which plumbing fixtures were in use. But they needed more data to make their program accurate enough for consumer use.
This was my first time hands-on with machine learning and reinforcement learning. I partnered deeply with the data scientists on the project to understand the details of how they were training their models, compressing and sending data, and how we could use reinforcement to refine results over time.
What people say
I designed an app that would help people set up their devices, and in the process - gather and label the training data needed.
The problem is - people think and say that they'll do the best job possible during setup - but my observations showed that in fact they took shortcuts, forgot fixtures, and would actually make the data science muddled.
Firstly, people forgot the fixtures they had in their house. Paper and pen prototypes helped me get to the bottom of this problem quickly. As a contractor, I didn't have time to run bloated studies, so I did scrappy, fast customer feedback. Rather than complicate things, I had the app lead them through their house, add the fixtures they saw, then turn them on and off in a routine that allowed white-space between each usage.
What people do
Reliably, no one waited for their toilet to flush to move on. 100% of people did not wait. "That's good enough, they'd say. But it wasn't for the data sicence - we needed a zero point between fixtures. Adding a 30 second timer any time a toilet was flushed, helped us get the data right.
On this job, I learned about and accounted for model drift, and reinforcement learning. I began to understand the elements of AI and how they relate to design, going on to teach my methods to others for the next decade.
Eventually, data collected, Phyn was able to release a later version that is plug-and-play for an effortless consumer experience.
