In a recent interview with SafetyDetectives, Qiuyan Xu, Managing Director at Gravitate AI, shared insights into her journey from earning a Ph.D. in Statistics to her current role in helping companies build AI applications and data pipelines. Gravitate AI’s flagship services encompass designing, developing, and deploying AI-powered products and processes, with a focus on solution architectures and modern AI technology stacks. Qiuyan also emphasized the importance of startups understanding where AI can enhance their products or processes and making strategic “buy or build” decisions based on their unique needs and long-term goals. Furthermore, she discussed how AI and machine learning are transforming traditional industries like healthcare, finance, and retail, and provided tips on measuring the ROI of AI investments, emphasizing a milestone-oriented, tangible roadmap for AI development.
Thank you Qiuyan for your time today, can you talk about your journey and your current role at Gravitate AI?
My background is in mathematics. I came to the US to earn my PhD in Statistics from UC Davis and then started out as a data practitioner using machine learning and AI to help businesses. My initial role was in insurance doing a lot of predictive modeling (including the distribution and sales side like propensity models) where I successfully deployed marketing, pricing and risk management models for US Fortune 500 companies. After that time was the big data trend when everyone started using big data and data cloud tools. I later spent several years building up big data analytics capabilities for FinTech startups until founding Gravitate AI in 2019. In my current role, I help a lot of companies, including startups, build their AI applications and the data pipelines that support their applications. While I serve primarily as the business leader, I still do a lot of design and advise clients across different industries on creating an AI solution to help their businesses scale.
What are Gravitate AI’s flagship services?
Our flagship services include the design, development and putting AI-powered products and processes into production for our customers. We specialize in solution architectures, detail-oriented model building, ML operation, and the modern AI technical stack, including LLM, NLP, image analysis using deep learning models, and their deployment in scalable systems. We’re a full service AI shop – from gathering initial requirements to maintaining a customer’s AI applications, most of which are a part of their software. So we create a microservice for their software platforms (whether web or mobile applications) that can directly call these AI services and return the results through the AI algorithms. Part of this is also building the data pipeline and accumulating data assets because the majority of AI algorithms must be customized and trained on industry or business specific data and that’s a big component. We also ensure the data components (the pipelines, the processes) are designed correctly and are scalable and robust, which is crucial.
What are the key considerations for startups when embarking on an AI journey?
Number one, I think, is to clearly understand where AI can help in their products or processes. I like to think about, from a product perspective, how AI helps the software product, for example, become more user friendly and/or more effective. And for processes, a lot of times the AI component helps them to streamline and become more efficient. So really they need to understand where these AI applications can be used. Then depending on the stage of the startup, what is the most appropriate way to deploy them? You can spend tons of money to do very complicated AI applications or sometimes you can just call Open AI and then wrap stuff around. But that all depends on the stage of the startup, how to prioritize development along with other business and product needs, and how to optimize a budget allocation while making sure the roadmap not only hits short-term goals, but also keeps the long term in mind so they don’t waste time and have to completely rebuild something later.
Can you elaborate on the “buy or build” decision when it comes to AI solutions? What factors should companies consider?
When it comes to “buy or build” I think there are two dimensions that are really important. First, whatever you buy or build really has to fit what you need to help the business. It doesn’t make sense to use something that’s out there, even if it’s cheaper, but it doesn’t 100% fit your needs. So the first step is really figuring out what is needed, then assessing the options. If there’s something available within budget that checks all the boxes, it doesn’t exclude the possibility to buy, but it becomes about really evaluating the different options against budget considerations. From a strategic perspective, startup founders should think about if that AI component is going to be part of their core IP. There are situations when it’s not necessarily the case. They might have a product built for an industry that solely relies on other features where AI only plays a supporting role to make things a little better. So in that kind of situation, AI might not be a key component for the startup. It may improve efficiency, make a product more usable but compared to the core product of what they built, the AI may not be mission critical and so that IP component might be less important. In that case, of course, ‘buying’ is completely fine but if it’s the other way around, the AI component in the long term together with the data component, can be a huge asset in which case the ‘build’ option definitely provides more flexibility to maximize and achieve IP value.
How are AI and Machine Learning transforming traditional industries like healthcare, finance, and retail?
In my opinion, there’s different levels of transformation. On one level there are a lot of key breakthroughs enabling something innovative that was not possible before. For example, in health care, a lot of drug discoveries require labs, clinical trials and huge amounts of capital and resources, and now we are seeing trends that AI might be able to take on some of those roles, potentially reducing the need for lab space, or human subjects. And while we’re not there yet, that’s the kind of transformation that can be a huge game changer. AI in that sense is becoming a discovery tool helping us make new and/or quicker discoveries that used to take a lot more time and effort. A different level of transformation is similar to when the internet or smartphone applications started, where there were select, smaller groups who were able to use them but eventually it became a necessary utility completely integrated into everyone’s daily lives. Similarly, I think AI and ML will become a common necessity across all industries where it transitions from “it’s good to have” to “you need to have” for more efficient services and products.
Can you provide some tips on how companies can measure the ROI of their AI investments?
Measuring the ROI of AI can be tricky, but I believe some AI components are a bit more tangible and easier to measure. For the AI component that is more product oriented, there’s a lot of research and analysis behind product performance, so if the AI component is part of a product, this can be measured and analyzed like any other feature and therefore easier to calculate ROI from that perspective. The difficult part is when companies want to spend their entire investment to create the kind of IP I spoke about earlier, and that gets into long-term ROI which is always more difficult to determine. My recommended approach is to make sure you have a long-term vision with a clearly defined, ultimate goal based on a very milestone oriented, tangible roadmap for AI development with smaller goals to achieve. Instead of spending a multi-million dollar investment on what you think is going to happen five years later, it’s about breaking down that investment using the agile product approach, applying AI, and being able to make it more tangible and measurable. I think that kind of dual approach is very important.There can be challenges, certainly, as we’re still determining how best to test and measure, and there’s still a lot of open questions, but the more it’s put into practice, the more best practices will be established.