How Haluk Demirkan is using ML-powered forecasts to get the right devices to the right place at the right time
For years, Amazon has been at the forefront of machine learning and data science. At the same time, the company has pioneered the large-scale automation of processes at all levels of its supply chains. But in its fast-moving commercial world, the constant challenge is to integrate these complementary fields to create processes that optimize the delivery of customer value.
In 2021, Haluk Demirkan joined Amazon to boost the company’s efforts for that kind of integration. As the senior manager of Devices Demand Science, Demirkan is building and leading a team dedicated to using cutting-edge data analytics, machine learning, and process optimization — among other techniques — to transform sales predictions for Amazon devices, from the Kindle, the Echo family of devices, and Fire Tablets, to Fire TVs, and Ring Video Doorbell — an enormously important task.
To understand why the company’s Devices organization is excited about Demirkan’s arrival, a little background is required. Demirkan first came to the US from Turkey in 1991 on a three-month language course to improve his English. Three decades later, he hasn’t left. Instead, he has built a pioneering transdisciplinary career at the intersection of data science, service science, smart machines, and industry.
Bridging industry and academia
He earned a master’s in industrial and systems engineering and in 2002 completed a dual-degree PhD in information systems and operations management at the University of Florida. He gained these qualifications while simultaneously working full-time for AT&T Bell Labs (as it was then known) and Citibank in data analytics, process engineering, and price and supply chain optimization.
By the time he earned his PhD, Demirkan had already spent 11 years in industry, so in 2002 he decided to give full-time academia a try. He joined Arizona State University as an assistant professor, primarily teaching information systems, analytics and supply chain management. While at ASU, Demirkan co-edited two seminal research books in the emerging fields of service science and systems, and its industry-based implementation.
In 2013, he moved to the University of Washington-Tacoma as an associate professor of service innovation and business analytics, tasked with expanding research and education programs in business analytics. He eventually became founding director of the Milgard School’s Center for Business Analytics and Master of Science Business Analytics, and the founding assistant dean of the Analytics Innovations Hub.
So far, he has published five books and more than 170 research articles. In 2021, Demirkan’s work and community building were recognized by the university with an award for both Distinguished Research and Community Engagement.
Throughout this academic stretch of his career, Demirkan maintained strong links with industry, developing data science, engineering and smart analytics solutions for dozens of leading companies, including IBM, GE, Cisco, HP, Intel, Bank of America, and Mayo Clinic. With AWS Academy Educator Accreditation, he still teaches a data analytics course at the University of Washington-Tacoma on weekends, guiding his students in developing AI and data-analytics-based solutions to novel business problems.
Researching “big problems”
So how did he end up full-time at Amazon?
“I’d come close to one of the highest positions in the academic career path,” says Demirkan. “The next step would have been to become a dean somewhere. But I didn’t want to be dean for near term: I prefer working on research for big problems.”
So, after almost two decades in academia, he decided to take a sabbatical. However, his restless nature meant he couldn’t be idle, so Demirkan applied to become an Amazon Scholar, a flexible program designed for academics who want to tackle large-scale technical challenges.
His plan: work one day a week during his sabbatical. During the interview with Amazon, however, it became clear that some of Amazon’s big business challenges dovetailed with Demirkan’s skillset so strongly — and offered him the opportunity to make such a big impact — that he decided to join the company full time by taking a leave from UW.
“My wife was like, ‘This is not a sabbatical!’”, recalls Demirkan.
The power of demand prediction
In his new Seattle-based role, Demirkan has two primary areas of business focus. The first is in making sales demand predictions for most Amazon devices. His team produces sales predictions for the majority of device types, globally, in which Amazon has a presence, and for every day from now until a year in the future.
To do this, Demirkan’s team ingests device sales data to train machine learning algorithms to generate increasingly accurate sales forecasts. Specifically, the team is employing advanced time-series forecasting methods, such as Random Forest, XGBoost, and Ridge Regression.
“Based on that predicted demand, another Amazon team ships just the right amount of product to the right locations, so it’s where it needs to be just as a customer chooses to make that purchase,” says Demirkan. “Basically, our goal is to get the right amount of devices to the right place at the right time to better meet our customers’ needs. That’s our goal.”
The forecasts Demirkan’s team creates do more than mere prediction. The team can also, for example, make projections for sales based on varying promotional prices.
“How many customers in London are going to buy a Kindle on April 21, if the price is x? The forecast assists our executive management teams to make decisions around how many units to manufacture, how many to ship, and when to ship,” says Demirkan.
By providing the company’s supply chain with increasingly accurate demand forecasts, Amazon simultaneously reduces delivery times and supply chain costs, helping the company keep prices low, while increasing customer responsiveness.
Demirkan is also developing a comprehensive, science-driven forecasting model called “Intelligent Demand Plan”. It will combine a wider range of inputs, including product cannibalization, macro-economic factors, traffic, social media and lots more, to sense demand and customer preferences with greater nuance, and to gain early insight into emerging market trends.
Automation and optimization
Demirkan’s second area of business focus: process and task automation and optimization, which utilizes his expertise in AI, process engineering and supply chain management. He and his team are analyzing the forecasting processes in Amazon’s devices group, identifying opportunities for improvement.
“Amazon is one of the fastest-moving companies I have ever seen,” he says. “Everything is about providing the best possible service to customers, and innovation is happening so quickly here that processes designed as recently as six months ago may no longer be optimal.”
This is one of his research passions: machine-assisted cognition, also known as intelligence augmentation with artificial intelligence.
“Computers are already our assistants today, of course. I want to take these computing technologies to a more advanced level, using machine learning to, for example, train computers to teach themselves to provide me with what I need to know to make better business decisions,” Demirkan explains. “By making processes more automated, efficient, and error proof, we humans have time to do more value-added tasks.”
Breaking research silos
Demirkan said he expects his team to grow in the next six months. “I have applied scientists, research scientists, and data scientists. It’s one of many fast-growing teams at Amazon,” he says.
Demirkan’s transdisciplinary expertise — that combination of deep research knowledge and broad applications experience — is something he will be infusing into his team’s culture. Many education systems, he argues, with their tendencies to silo students in particular domains of expertise, are producing a generation of people who can find it hard to adapt to the wider commercial world.
“With the digital transformation of companies in every industry, the coming generations of employees need to be more ‘T-shaped’ — innovators with not only a depth of expertise but also a broad, integrated understanding of other disciplines, such as management, engineering, and social sciences,” he says.
Last year, Demirkan’s insights into the changing nature of the high-tech workplace were sought by the US Department of Labor, Employment and Training Administration, when he was invited as an occupation expert to provide guidance on the definition of the occupation Operations Research Analysts.
“This role is about the combination of operations management, IT, data science, and machine learning — a very multidisciplinary, new occupation,” says Demirkan. “I’m hoping that in the future, academia will have more programs geared to preparing people for these crucial kinds of roles.”
With new roles in mind, how has Demirkan enjoyed his work with Amazon?
“A friend of mine said to me: ‘Haluk, you are going 35 miles an hour in academia, and now you are switching to 200 miles an hour?!’,” he says. “I love being a professor and making a difference in students’ lives, but I am relishing being back in industry because in 19 years, things have changed. I’m absorbing so much, and I can take this updated knowledge back to my classes when I teach on the weekend.”
From his professional perspective, Demirkan sees more clearly than most the gaps between academic education and the expectations of professional workplaces. “Globally, we have an ongoing mismatch problem. With my experience with Amazon, I can do my bit to close this gap,” he says.
Giving back
Doing his bit is central to Demirkan’s ethos. When Covid-19 struck in 2020, and hospitals all over the planet were suddenly critically overloaded, Demirkan was contacted by Virginia Mason Franciscan Health, one of the largest healthcare service and hospital providers in Washington state. They wanted his help to optimize their hospital bed allocations, among other things.
“We met online every week, trying to predict demand and capacity, which patients to move to other hospital facilities, looking at doctors’ and nurses’ scheduling — everything.” The urgency of the situation meant Demirkan was more than a volunteer advisor. “I was writing machine learning scripts, literally writing the code myself, to exploit the data quickly being gathered by the hospitals. I was proud to be involved in that work, because it was the first time I was able to make that sort of critical difference in people’s lives.”
In terms of making a difference at Amazon, Demirkan says: “My goal is to build an ecosystem in which technology is doing the labor-intensive tasks, freeing my team to do more smart work and value-added tasks. That’s my idea of success.” That, and taking Amazon’s device-sales forecasting to a whole new level. “I want my team forecasting at a comprehensive, granular level,” says Demirkan. “I want to say with unprecedented accuracy that in this location, this device — in this color, size, and detail — will sell x units on this day.”
But Demirkan also sees additional potential in developing approaches that go beyond traditional forecasting. “I want to develop machine learning and data analytics that can discern what it is that customers really want and expect from Amazon devices; to generate insights powerful enough to actually impact the design decisions for brand new products and services.”
