In the Author Spotlight series, TDS editors talk to members of our community about their career path in Data Science and AI, their writing, and their sources of inspiration. Today, we are happy to share our conversations with them Shuai guo.
Shuai is an industrial AI researcher who works with physics, data, and machine learning to solve real-world problems in Engineering, security, and intelligent systems. He holds a PhD at the intersection of computational psychology and machine learning. His work covers a variety of topics, including anomaly detection, digital twins, physics learning, and LLM/agentic.
– Yours A piece of langgraph He walks the reader through the process of creating a deep research agent. When you actually tried the ending, what surprised you the most, and what would you do differently next time?
I can say that what surprised me the most was how a serious research owner could make mistakes when running at the end. That “issued the question. There are two main problems that I remember well. First, from time to time, the agent begins to mix the findings with what he remembers from the previous training. This is not good, since I only want llms that include information and target information gaps, while relying entirely on web searches to find an answer.
Another issue that always gives me a headache is data contamination, which is when a search returns the same items but the model behaves like it's exactly what you asked for. For example, I once tested a deep research agent by researching a specific bug report (say, issue # 4521 of the code), and the search will return content related to their symptoms as all the same problem.
Apart from these two main challenges, I also found challenges in handling conflicting information and determining adequacy to complete an in-depth study. None of those problems can be solved by simply adding more search results or more functionality.
An important observation for me is that Guardrails is as important, if not more, than the agent build, if we want to go beyond “just a demo” and build a system that works in production. I think the mindset of “Test-advent Development” fits well here: Define what “good” looks like before you build. Next time, I would start by defining specific rules, then build the agent's architecture around those constraints.
You wrote that Analytical AI (SQL / Bi + Classical ML) is not going away because agents are hot. If you were designing a modern data stack today, what task would you assign to agents and what would you keep in the analysis pipeline?
Analytical Ai is revolutionary and statistically accurate. LLM-based agents, on the other hand, are good at digesting raw context, interpreting results, and interacting with people. For the division of labor between Analytical AI and Agentic AI, I would say that if the work was more correlated, I would make a mistake in AI; But if it is more intended, e.g.
We can look at the concrete problem of building a customer forecasting system. At a high level, it usually involves two steps: identifying vulnerable customers, and acting on them. According to the first step of flagging vulnerable customers, I would rely on APAYTICAL AI to Engineer Informative Pictures, train gradient amplification models on historical behavioral data, and then use the trained models to be intelligent. In addition, I also conducted a categorical analysis to find the most important scores in predicting the interpretation. All the steps are accurate and refreshing, and there are a ton of best practices for getting accurate and reliable results.
But then comes the fun part: What do you actually do with those predictions? This is where LLM-based agents can take over. They can distribute personalized maintenance emails by drawing on a customer's history, perhaps suggesting relevant product features they haven't tried yet, and adjusting the tone based on how their previous support tickets went. There are no statistics here. Just talking about content wise.
What is one skill you invested in at our startup that is now paying off as AI tools become more powerful?
Programs think.
To me, systems thinking is basically asking to decompose systems into parts. How do the different parts talk to each other? What are hand points? Where are the answer holes? When I touch this onewhat else is changing?
I chose this at university. I majored in aerospace engineering with a focus on Aero-Engine design. The thing about jet engines is that everything affects everything, and studying it really helped me develop three habits: decomposing the system, and explaining the clean results.
It is true that AI tools are gaining more and more capable power, we have found better process assistants, more efficient pipelines, or llms that can manage a long context, but most of the development takes place at small risks. Instead of chasing the hottest tool and trying to fit it somehow into my existing work, systems thinking helps me put the big picture front and center. With the LLM program, I always started by drawing things, decided to interact with the inputs / outputs between the elements, making sure that checks and ogades are included when the tools develop.
In fact, LLM application building systems remind me a lot of jet engine design: new technologies come and go, but the accomplice system compounds value.
Zooming in, what part of science or AI is changing the most right now, and what part is changing the fastest?
I think agent ai programming is definitely one of the hottest fields that is moving very fast. We see great demos (be it coding assistant or research assistant) every now and then. New open source frameworks that enable developers to better build their own multi-agent systems are also appearing regularly. All of this is exciting. But here's the thing: Are we forcing these complex systems a faster way than understanding how they actually behave?
This is where I see the gap: all the “Assurance” layers around those multitasking programs don't appear fast enough. To meet this challenge, we can (and probably will) manage those multi-agent systems like any other industrial system. In the manufacturing industry, it is common practice to adopt data-driven methods to facilitate system design, control, condition monitoring and error analysis. This same approach can benefit agent systems. For example, how, how do we use Bayesian efficiency to design multiple architectures? What about using ML anomaly detection to monitor the performance of agents and catch security threats?
The good news is the Momentum structure. We're seeing LLMS visualization platforms, evaluation frameworks, etc., and laying the groundwork for using those industries in industrial-grade, data-driven ways. I see a lot of opportunities in this area and that's what excites me: the opportunity to bring the complexity of industrial systems to Agentic AI and make those tools reliable and trustworthy.
To learn more about Shuai's work and stay up to date with his latest articles, you can follow him on TDS or LinkedIn.
