iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI that performs well in a controlled experiment still has to cope with unfamiliar inputs, changing conditions, limited resources, and the cost of keeping it running. In a 30 September 2026 interview with The AI Journal, software engineer and researcher Prajval Mohan connects that practical challenge to his work in reinforcement learning, computer vision, digital pathology, and scalable software systems. His account offers a useful way to think about deployment: assess not only what a model can do under ideal conditions, but also whether people can reproduce, operate, and maintain it in the conditions that matter.
What connects Mohan’s work across different AI fields?
Mohan says he looks for problems that are practically important, technically meaningful, and capable of producing something useful. That outlook connects work as varied as path planning, roadside safety, digital pathology, and systems for the mortgage industry. As he puts it, “What connects these areas is a focus on building practical systems where reliability, efficiency, and real-world constraints matter.” Read the interview in The AI Journal.
The through-line is not a single algorithm or application. It is the effort to make a technical result work in a setting where inputs, resources, and operating conditions may not match the assumptions of a controlled experiment.
Why can reinforcement learning be difficult to deploy?
In the interview, Mohan describes exploration as a deployment trade-off. An agent may need to try actions to learn, but a poor action can be costly or unsafe in a real environment. Restricting exploration can reduce that risk while also limiting the quality of the solution the agent discovers.
#1 Best Overall
He also points to changing environments, limited computation or information, and tight decision-time budgets. A policy that succeeds during training may become unstable when conditions shift or an unexpected event occurs. “A policy that works during training may become unstable when the environment changes or something unexpected occurs,” he says.
These are Mohan’s engineering concerns, not a claim that every reinforcement-learning application faces the same risks or that one method resolves them. The interview mentions a path-planning effort called “Iterative SARSA,” but does not provide its publication, implementation, algorithmic details, or evaluation results. The name alone is not enough to assess its novelty or performance.
What does the Slideflow work show about computer vision in practice?
Digital pathology makes the practical demands of computer vision especially visible: systems must process large medical images, and users may need to understand how models behave as well as whether they produce predictions. Mohan says his work with Slideflow involved capabilities such as model ensembles and out-of-distribution detection, and mentions deep ensembles, hyper-deep ensembles, and adversarial training. Those specific contributions are his account in the interview.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesThe project’s repository describes Slideflow as open-source deep-learning software for digital pathology, with image-processing tools, uncertainty quantification, and explainability. See the Slideflow project repository. A 2024 paper co-authored by Mohan describes the library and its whole-slide visualization interface. Its authors report that tile extraction at 40x magnification took 2.5 seconds per slide; that is a paper-reported result, not a guarantee for other hardware or workloads. Read the 2024 Slideflow paper.
The repository and paper establish Mohan’s association with Slideflow and describe the framework’s capabilities. They do not, in the material available here, independently verify every feature contribution the interview attributes specifically to him.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should AI research results account for deployment?
Mohan recommends distinguishing between what a system achieves under ideal conditions and what can realistically be maintained after cost, scale, and ongoing upkeep are considered. “I think it would help to report two kinds of results,” he says. This is his proposed framing in the interview, not a standardized reporting rule.
| Result to report | What it helps readers understand |
|---|---|
| Ideal-condition performance | What the system achieves under the experimental conditions chosen by the researchers. |
| Practical, maintainable performance | What remains achievable when operational cost, scale, and long-term upkeep are taken into account. |
For Mohan, deployment-minded design also means asking whether others can reproduce a result, whether it can be implemented reliably, and how it responds to unusual inputs, system failures, heavier workloads, or changing conditions. These questions make the boundary between a promising experiment and a dependable system easier to see.
Recommended Free Tools
Quick Recap
What readers can take from Mohan’s perspective
- Judge the operating conditions, not just the headline result. Consider what happens when inputs or the environment depart from the expected case.
- Make risk and resources visible. In reinforcement learning, exploration may have real costs; computation, information, and time can also constrain what is feasible.
- Treat reproducibility and maintenance as engineering questions. A result is more useful when another team can implement it and understand the burden of keeping it reliable.
- Separate evidence from attribution. The Slideflow project and paper document the framework; Mohan’s interview provides the account of his particular contributions.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
