AI research often looks impressive in a controlled environment. The harder question is what happens when those systems have to work with messy data, enormous datasets, unpredictable conditions, and real-world constraints. That is the problem Prajval Mohan has spent much of his career exploring.
A software engineer and researcher specializing in artificial intelligence, machine learning, computer vision, and scalable software systems, Mohan has worked across areas ranging from digital pathology to reinforcement learning and trajectory planning. His research includes contributions to Slideflow, an open-source deep learning framework for digital pathology, as well as the development of Iterative SARSA, a reinforcement learning approach designed for complex environments.
His work sits at the point where AI theory meets engineering reality: how to build sophisticated models, how to scale them, and how to make them useful outside the research setting.
In this interview, Mohan discusses what it takes to move AI from experimentation to real-world deployment, the challenges of building reliable intelligent systems, and where he sees the next wave of progress in computer vision, reinforcement learning, and applied AI.
Your work spans reinforcement learning, computer vision, digital pathology, and scalable software systems. What connects these areas for you, and how do you decide which technical problems are worth pursuing?
My interest in research developed from my tendency to notice problems and look for practical ways to solve them. With a strong computer science background, I decided not to limit myself to one area. I began looking across domains for real problems that could benefit from my technical knowledge. I examined limitations in path planning, which led to my work on Iterative SARSA. I later observed the dependence on human attention in roadside safety, which led me to explore helmet and tire-defect detection. My work with Slideflow gave me the opportunity to apply computer vision to digital pathology, where technology can help researchers address challenges such as limited access to specialized expertise. Today, I apply the same approach as I design scalable systems for the mortgage industry. What connects these areas is a focus on building practical systems where reliability, efficiency, and real-world constraints matter. I pursue problems that have practical importance, present a meaningful technical challenge, and offer the possibility of creating something people can actually use.
When applying reinforcement learning to complex environments, what are the hardest problems around exploration, convergence, stability, and real-world constraints?
Reinforcement learning can perform well in a controlled environment, but applying it in the real world creates a different set of challenges. An agent has to try unfamiliar actions to learn, but in the real world, a bad decision can be costly or unsafe. At the same time, restricting exploration too much may cause it to settle on a weaker solution. Convergence is therefore not enough by itself. A policy that works during training may become unstable when the environment changes or something unexpected occurs. There are also limits on computation, available information, and how quickly a decision must be made. The difficult part is dealing with all of these issues together because improving one can often make another worse.
Your work with Slideflow involved deep learning and digital pathology. What did that experience teach you about building computer vision systems for domains where accuracy, interpretability, and reliability are especially important?
Accuracy and reliability are important for any AI or deep learning system, but in medicine and healthcare, they carry much greater weight. My work with Slideflow was closely tied to that responsibility. I designed and implemented capabilities that were not previously available in Slideflow, including functionality for model ensembles and out-of-distribution detection. My work included deep ensembles, hyper-deep ensembles, and adversarial training, helping the framework evaluate uncertainty and recognize patterns outside the training datasets. These contributions were integrated into Slideflow and the functions I contributed remain part of the framework. I also co-authored the Slideflow paper in BMC Bioinformatics, which has since been cited more than 100 times. That experience changed how I think about accuracy. A prediction may look accurate, but in a field such as healthcare, users also need to know how confident the model is and whether its output should be trusted. It also taught me to be extremely cautious as a developer because even a small oversight can have serious consequences for the end user. I have carried that mindset into every project I have worked on since, especially when designing systems that need to remain reliable as they scale.
You have worked across both academic research and software engineering. How has your engineering experience influenced the way you design, test, and evaluate AI research?
My engineering background has helped me approach research problems from a practical perspective. I do not stop at whether an idea works in theory. I also think, “Can someone else reproduce it?”, “Can it be implemented reliably?” and “What will happen as its use grows?”. Engineering has also taught me to ask what happens outside the ideal test case. How does the system respond to an unusual input? I have learned that some problems do not appear during controlled testing. They may only surface when the system fails, handles a heavier workload, or operates under changing conditions. As a result, I evaluate a system not only by its experimental accuracy, but also by how consistently it behaves when the conditions are no longer ideal.
Many AI systems perform well in controlled experiments but become difficult to scale or maintain in real-world settings. What do you think researchers should consider earlier when designing systems intended for practical deployment?
Academic research often focuses on achieving the highest possible accuracy. To accomplish this, experiments are usually conducted in controlled and carefully tuned environments. This is important for understanding a problem’s theoretical limits, but it can also become a bottleneck when research moves to practical applications. I think it would help to report two kinds of results. One would show what the method can achieve under ideal conditions. The other would show what can realistically be maintained once cost, scale, and long-term upkeep are factored in. This would give implementers a clearer understanding of the performance they can realistically expect. The basic research design may not need to change significantly, but how the system is tuned and evaluated can strongly affect whether it succeeds in the real world.
Looking across your work in reinforcement learning, computer vision, and scalable AI systems, which research questions do you think are still being underestimated, and where do you see the most promising opportunities for future work?
As AI transforms how people think and work, I believe the most promising opportunities are no longer purely theoretical. We also need to focus on what these advances can help us build and solve. AI has already made major advances in models and learning algorithms, but what I think is still underestimated is the work required to turn those advances into systems people can depend on. I have encountered this challenge across research and engineering: an approach that performs well experimentally still has to remain reliable under real-world constraints, changing conditions, and scale. I see opportunities in applying existing AI capabilities to problems that once seemed impractical, whether in autonomous systems, healthcare, or scientific discovery. There is still room for theoretical progress, but I believe some of AI’s most useful advances will come from finding better ways to turn existing capabilities into dependable systems.
