AI is reshaping our world at an unprecedented pace, yet the path from research breakthroughs to industry-ready solutions doesn’t happen overnight. Turning technical innovations into practical tools takes more than cutting-edge algorithms. It also requires a clear understanding of industry needs—and close collaboration across disciplines.
At Microsoft Research Asia, a team of researchers is working to bring AI into real-world applications. Among them is Xiaofan Gui—a scientist who combines abstract thinking with a pragmatic mindset. She and her colleagues are exploring new ways to expand what AI can do outside the lab.
This article examines Gui’s unique approach to translating research insights into practical, real-world solutions.
The true value of technology
After getting her degree in computer science, Gui joined a startup, where she developed a platform for trading used books on college campuses. This experience helped her realize that the true value of innovative technology lies in solving real-world problems—and that implementing it is often challenging. To further develop her skills, she attended the School of Software and Microelectronics at Peking University, where she received a master’s degree in software engineering.
There, Gui came across an opportunity that changed the course of her career: a class involving close collaboration with Microsoft Research Asia. Through this class, she discovered that the organization not only conducts fundamental research but also applies innovative technology to solve real problems—a mission that deeply resonated with her. She secured an internship at Microsoft Research Asia, working with colleagues to turn cutting-edge research into practical tools that meet real user needs.
Her first project involved developing an English learning platform that applied Microsoft Research Asia’s algorithms to real-world contexts. The experience reinforced her impression that the organization is driven by a commitment to solving real-world challenges through technology—an outlook that strengthened her determination contribute to its mission. “Microsoft Research Asia has strong technical capabilities, and its diverse and inclusive culture creates a comfortable and friendly research environment,” she said.
After earning her master’s degree, Gui joined the machine learning group and contributed to multiple industry collaborations. These include predicting the health of Nissan car batteries with machine learning, exploring effective strategies for global carbon budgets using AI, and helping telecom companies detect malicious websites and lateral movements (a technique commonly used by cyber attackers) through prediction models. For the past three years, Gui and her colleagues have remained committed to applying technology to real-world problems, advancing the integration of AI across industries.
The first step in integrating AI and industry
Implementing AI in industry involves more than simply building a model. The first step consists of closely engaging with real-world use cases, translating industry needs into trainable algorithmic tasks, and delivering tangible value.
“Each industry scenario is like a unique puzzle, requiring us to first find the underlying patterns and then design or choose the most suitable algorithm,” said Gui. As an applied scientist, her core work involves converting problems into practical algorithmic models—and ensuring those models remain understandable and usable in practice.
For example, for a project involving ocean carbon sink research—a collaboration between Tsinghua University and the French Environmental Sciences Laboratory—Gui and her colleagues combined marine biogeochemical knowledge and data-driven models to design a new machine learning simulator.
Previously, ocean carbon budgets were handled in two ways. One relied on numerical simulation, which was slow to update and couldn’t easily incorporate new data. The other used data collected from ship sensors and satellite monitoring to build machine learning models. But because this data came from specific locations at specific times, the models had to generalize from limited snapshots to global conditions—often at the cost of accuracy.
To address these challenges, Gui reframed the oceancarbon budget problem as one that required integrating different types of data. She combined numerical simulation and survey data to build a model that could dynamically adapt to various marine regions. Finally, she integrated marine biogeochemical knowledge to optimize the model’s parameters and structure, aligning it more closely with real environmental changes.
“The application of AI doesn’t solely rely on precise algorithms—it also requires deep domain knowledge,” Gui explained. “Only when algorithms are closely aligned with real-world tasks can they truly bring value.”
Turning data processing into a craft
High-quality data is essential for building effective algorithms. Gui is highly attuned to this fact and finds satisfaction in work that others might consider tedious. She also sees data processing not just as fixing errors and filling in missing values, but as a craft that requires precision and care.
She separates data processing into the following steps: standardization, anomaly detection, filling, correlation analysis, and feature extraction—all aimed at uncovering the value in raw data.
The project to predict Nissan car battery health was an ideal opportunity to apply Gui’s approach. Because battery charging and discharging cycle data is both complex and sparse, using standard methods to streamline or simplify the data risked losing important information.
Gui’s method involved first standardizing the data format to ensure consistency across the analysis. Next, she identified anomalies—for example, a power outage during a cycle would be considered abnormal. She consulted industry experts to confirm whether these anomalies should be retained or removed to help ensure data accuracy. When the data was incomplete or too sparse, Gui supplemented it with public datasets or historical and neighboring battery data.
Once the data was cleaned, she analyzed it to identify correlations and extract valuable features. To ensure the scientific soundness of her method, she reviewed relevant literature to understand the industry consensus on key features.
For example, during battery charge and discharge cycles, voltage and capacity changes over time. A significant drop in capacity within the first 100 cycles may indicate a short battery lifespan, while a more gradual decline may suggest a longer life.
By processing the data this way, researchers at Microsoft Research Asia could design features that can predict battery health at 800 cycles using data from only the first 50—making Nissan’s battery monitoring and management more efficient and intelligent.
Another example of Gui’s problem-solving approach involves the detection of malicious websites by telecommunications companies. These websites come in many categories, and the data involved is both large and varied. Phishing websites, for instance, frequently change domain names and use redirects. Hacked, or defaced, domain names can appear normal but often embed malicious ads or alter parts of the content—making them difficult for traditional methods to detect.
To address this, Gui adopted a classification-based approach. For phishing sites, she developed a detection algorithm that analyzes the relationship between website content and domain names. By extracting trademark information, tracing affiliated companies, and comparing content with domain data, the system could more accurately identify phishing websites. For defaced websites, she compared current site content with cached versions and incorporated data from a “hacker attack” corpus to identify suspicious alterations.
For Gui, the seemingly monotonous task of data processing is, in fact, a process of uncovering hidden patterns within an industry. “When I delve into the chemical reaction mechanisms of batteries to understand the causes of degradation or study hacker techniques to detect malicious websites, I gain a sense of satisfaction from learning new things and exploring root causes. Every time I overcome a seemingly tedious detail, my sense of accomplishment doubles,” she said.
Cross-domain collaboration fosters infinite possibilities
Another key element in advancing the implementation of AI is working side by side with experts from different industries and disciplines. “To truly realize the value of AI, multidisciplinary collaboration is indispensable,” said Gui. “Many difficult problems span multiple domains, and only by combining perspectives and ideas from various fields can we find optimal solutions. The beauty of cooperation lies in being exposed to the diversity of perspectives across domains.”
Cross-disciplinary research often comes with communication challenges. But in Gui’s view, curiosity can bridge that gap. “If you’re interested in a field, take the first step—reach out to professionals, read the relevant literature, and learn to understand and respect their ideas,” she said. Whether through broad exploration or in-depth research, continuous dialogue fuels both professional and personal growth.
“Cross-disciplinary research demands both courage and curiosity,” she added. “If you’re willing to speak up and step into unfamiliar territory, you’ll discover a world that’s broader than you imagined—and full of new possibilities.”
Artificial intelligence