Mark Chung was a chip guy. Not a software developer, certainly not a power systems engineer. He spent nearly 15 years in the semiconductor industry since getting his master’s and bachelor’s degrees in electrical engineering at Stanford in 1999. He’d been an engineer at AMD for nearly six years designing chips like the Athlon and Opteron, then at startup PA Semi working on microprocessors that, he anticipated, would go into Apple computers (Apple later purchased the company and the designs would end up in iPhones).
In 2008 Chung was a principal engineer at RMI, a company that later merged with Netlogic and was acquired by Broadcom. One month that year—a month he and his family had mostly spent out of town, he received a surprisingly large electrical bill: $560, when his typical bill was around $100. He called the local electric company, and a representative assured him that his smart meter was working just fine, and the bill was correct. At work that week, he got into what he called “an engineer debate” with colleague Jonathan Chu over what could have caused the large bill and how to trace the source of the problem. That weekend he and Chu purchased two Kill-A-Watt meters—inexpensive plug-in devices designed to monitor the electricity use of appliances—and the two went from room to room around Chung’s house, spot-checking various appliances and gadgets.
“We didn’t see anything wrong,” he says.
Perhaps, the two reasoned, the problem was intermittent, and could only be identified with longer-term monitoring. So they hacked the meters to add WiFi chipsets and send data onto the home’s wireless network, writing software on an old Dell Inspiron to gather and display the data.
No anomalies emerged. And the next month the electric bill was just as high.
“That’s when we had an epiphany,” Chung says. “We realized we were doing this a stupid way; the smart way would be to just look at one point at all the electricity and unpack it to figure out where it is going.”
He did a quick search of the literature, and, he says, stumbled on a paper by Zico Kolter explaining an experimental theory involving tracking the electricity use of multiple appliances in a home by monitoring the energy data from a smart meter. Chung and Chu posited that a similar approach could be applied to tracking multiple systems on a single circuit at higher fidelity by applying packet inspection techniques to the signals coming from that circuit.
So Chung and Chu bought an old HP oscilloscope from Craigslist and mounted it on the main electrical panel feeding Chung’s house. They then set about rewriting an algorithm they’d been researching at work, one designed to accelerate packet separation in chips, to apply it to this problem.
Then, he says, something amazing happened. It worked… sort of. They realized they could see each individual device drawing power in Chung’s house at the moment it turned on.
“At first, we didn’t know what the devices were, but we see what they are doing. The hairdryer comes on for 5 minutes at 8 a.m. The Xbox in my brother-in-law’s room is on for 4 hours from 10p to 3am. We thought we could teach the system to label these things.”
And, it turned out, the pool pump, programmed to come on every day and run for 12 hours, was using a ridiculous amount of power—some 4 kW, which turned out to be about 10 times as much as it was supposed to. They had found the cause of the high electric bills.
He bought and installed a new pump. Problem solved, and Chung didn’t think much more about electricity use, he says, until 2010, when his first child was born.
“I had a sudden change in the time frame in which I was thinking about problems,” he said, “to thinking about another generation and a more distant future. It was transformational.”
And it brought him back to his hunt for his home electricity waster in 2008. “Someone who didn’t have the ability to figure out the problem,” he said, “would not just be paying for wasted electricity, they would be polluting the environment and wasting my child’s resources.”
He took a look at the market—and saw that, since 2008, a few companies had come out with electricity analyzers that installed in the circuit breaker box. However, they all had to be wired into the circuits by an electrician, and, he says, were “pretty crummy in terms of capability.”
“I thought if I pulled Jon [Chu] and a couple of other friends together, we could start a company and make something better.”
“I couldn’t convince them at that point to leave what were fairly cushy jobs at NetLogic,” he says. (At that point Broadcom had just bought the entire company for nearly $4 billion.) He decided to go it alone, and at the end of 2011, he quit his job. “I had to convince my wife to give me a year (it took nearly 2) to get the business off the ground and to let me invest some of money I’d made in the PA Semi and NetLogic acquisitions,” he says.
He started talking to potential customers, generally the people who managed hotels, airports, and other large businesses. “You have to go for people who can afford it first, when you are going for a new market,” Chung said.
It turned out these folks used expensive building management systems that generally just tracked cooling and heating, and that they did little with the data because it didn’t tell them much; they could indeed use a tool that told them more.
Armed with this kind of market research data, and some early pilots he conducted at the end of 2012, Chung convinced Chu, along with his brother Thomas Chung, to jump in. They named the company Verdigris, after the name of the green pigment that forms when copper is exposed to the environment. Then they spent three intensive months working on both the technology and the business development under the umbrella of the Stanford StartX accelerator, then moved to accelerator Founder.org as part of its inaugural class.
Along the way they connected with NASA, where researchers are interested in energy management systems to help develop technology for future extraterrestrial buildings, and NASA offered the company inexpensive office space along with an opportunity to use its Sustainability Base, a building designed to be as smart and green as possible using today’s technology, as a testbed.
“That,” Mark Chung said, “was perfect—we needed a way to collect detailed label sets on equipment, and NASA had redundant energy management systems along with smart plugs on every outlet, all data we could use to train our AI system.”
Fast forward to today. Verdigris has developed a magnetic sensor that clips on the outside of the wires leading into the circuit breaker box. It works by sampling changes in the magnetic field around the wires at a rate of 8 kilohertz. These types of sensors already existed, but most operated at lower frequencies and are larger—too large to fit on every wire in the tight space inside a typical circuit breaker box.
“We shrunk the sensors and got high fidelity signals,” Chung says, “by figuring out novel ways to fit more powerful sensors into tight spaces.” Some of that had to do with the arrangement of the magnetic material, he says, but can’t say more until patents are filed.
Verdigris also had to develop algorithms to decode the signals. It did so by training a deep learning system, essentially, turning things on and off while the system observed the changes in magnetic fingerprints. The system ships today with some basic labels in place, like refrigeration, lights, pumps, motors. If the user wants more specific information (GE microwave here, Samsung refrigerator there) they need to teach the AI themselves by labeling those devices when they are detected through the app.
“We can go down to five-watt devices,” Chung said, though he admits the system has a hard time telling the difference between an iPhone and iPad. “We could theoretically tell you if a computer is idle or being used if it’s on the charger.”
Beta systems rolled out to customers in 2014. Its production system started shipping late last year at US $3300 for a 42-sensor system; the company also charges customers $49 or $69 a month, depending on the level of data provided, for the cloud subscription. At this point, it is still targeting the business market, but the company does already have a handful of private residential customers in Silicon Valley.
One of its early customers used the product to solve a mystery much like the one that inspired its creation: Somewhere in the company somebody was doing something, likely bringing in a space heater, that was tripping a circuit breaker regularly. But nobody could find the culprit. “We said ‘we think we can,’ and we did,” says Chung.
Generally, he thought users would be using the data to improve energy efficiency. It turns out that, for now, says Chung, “our customers don’t seem to have time to care about that—they care more about whether a hotel room is going to be too hot or cold or whether machinery is about to fail. “We can detect when motors are faulty, for example, that a bearing has failed in a motor, or a rotor bar in a condenser pump has broken—we can pick up changes in the physical nature of devices from changes in their signatures.”
Verdigris has more than $16 million in funding, including about $1.5 million in startup funds from its founders and significant investments from manufacturing firm Jabil and Verizon. It has 28 employees, virtually all engineers, and is still hiring. Right now, it is focusing on getting this first product out to business customers.
But a consumer product is coming—“in a few short years,” Chung promises.
“Our goal is to eventually get down our hardware kit to $100 or less,” says Chung. “Then we can connect every residence, parking lot, pool house, connecting the dots between demand and loads on the grid, and help optimize the use of energy resources and save the planet.”
Updated 4 May 2017