In The Arena by TechArena - AI in Biomedical Engineering: Maya Kalyan on Diagnostics

Episode Date: June 4, 2026

This episode of Data Insights features Allyson Klein and Jeniece Wnorowski in conversation with Maya Kalyan, Staff Algorithms & AI Engineer. The discussion examines how AI is being operationalized in biomedical engineering—specifically in diagnostics, where improvements in accuracy, speed, cost, and scalability have direct clinical implications.

Welcome to Tech Arena, featuring authentic discussions between tech's leading innovators and our host, Alison Klein. Now, let's step into the arena. Welcome in the arena. My name's Allison Klein, and today is a Data Insights episode, so that means I am here with Janice Gerowski. Welcome back. Janice, how you doing? Hi, Allison. Thank you. I'm doing great, and it's awesome to be back, as always. So, Janice, we have a really exciting topic today. I want you to introduce it and our guest because this is a really special episode. Yeah, I think today is going to be a really special episode.

We have not done an episode like this ever. We've never talked about biomedical engineering and AI. And today we have Maya Calli Ann from Cerro Fisher, who's going to dive into some insights on how all of her work in biomedical engineering is affected by AI. So, Maya, welcome to the program. Thank you, Janice and Madison, for having me. I appreciate the invitation. It's great to be here.

So, Maya, I was so excited to talk to you. I know a little bit about Thermo Fisher. I know that you guys make things like electron microscopes and tons of different deep technologies that help in the life sciences arena. You're a data science and machine learning engineer. Can you just introduce a little bit about your background and about the background in biomedical engineering and how it plays with the industry?

Absolutely. I've been in the life science and biotic industry for, over 10 years now. My educational background is in biomedical engineering. I did my undergrad in SSL India before moving to the U.S. for my master's at the University of Michigan. What really drew me to biomedical engineering early on was that interdisciplinary mindset where you're trained to talk about how to bridge medicine, data, and engineering to solve real world healthcare problems. And that foundation is what led me to R&D roles for most of my career. And most recently, as you've mentioned,

I've transitioned into staff algorithms and AI engineered at Thermo Fisher, scientific, specifically in the molecular diagnostics space. What I do in this role is I'm collaborating with cross-functioning teams to explore how AI, including the emerging generative AI systems, can be used to advance our own research capabilities. and there's the next generation of diagnostic technologies. I'd say it's an ultra-tively exciting space to be in right now. Outside of work, I've also been spending time experimenting with AI

through research projects and hackathars. I think it's a great way to stay current with the new tours and technologies that are evolving so quickly. So, Maya, you work closely with research design and new product development for biomedical diagnostics. problems are you seeing in teams in this space? And what are you most focused on in solving right now? Yeah, that's a great question. There's a lot of innovation happening in diagnostics right now. When I think about problems that these innovations are focused on,

I would probably highlight four areas. The first one and foremost is accuracy and reliability. So diagnostic tests need to perform consistently across different environmental conditions like samples, labs. And so there's a lot of work going on into assay design and identifying quality issues to ensure reproducibility of results. Second, there is turnaround time because clinicians often need results quickly to make treatment decisions. So teens are actively working on ways to shorten time to results and streamline workflows in the lab. The third focus area I can think of is reducing cost. There's definitely a growing demand to make molecular testing more accessible. And innovations like multiplexing are helping with that by allowing multiple species to be detected in a single sample.

And for example, we have respiratory virus tests that can detect COVID and flu and RSV viruses all within the same test. So that reduces the reagent use and lowers your consumable cost, while also increasing throughput. And finally, another big focus right now is automation. So labs process, like huge volumes of samples, and there's always research going on on automating some of these complex steps in the workflow, like sample preparation,

so that there is more efficiency and scalability in the workflow. So I would say overall the goal is to build diagnostic systems that are not only faster, but more reliable, more often, affordable and easier for high throughput research and clinical labs to use. We talk a lot about AI on this program, and I think that one thing that I know from being involved in AI for the past decade is that life sciences has really been blazing a trail in terms of utilization of AI in all of the things that you just described and more. One thing that I wanted to ask you about is across the landscape of your professional and even external projects that

you're working on, where does AI fit in terms of being a critical path to a solution within life sciences and where does it potentially fall short? Yeah, that's a great question. I think generally speaking, AI tends to work well when you have large, well-structured data sets and a clear outcome you're trying to predict or detect. Some examples I can think of like patent recognition in biological data or quality monitoring in your experimental workflows. In those use cases, you know, foundational models can be trained with the data to uncover patterns that are normally too complex for traditional approaches. And one other area where we're seeing practical adoption is these domain-specific chatbots

or AI assistance. These systems can help researchers with experimental setup and their data analysis. or quickly looking out long user documentation to support quick troubleshooting. The benefit here really is that they can improve the user experience and reduce manual touchpoints by supporting more of these human-assisted workshows. Where AI tends to fall short currently is when data is not representative of the world variability or not labeled very well, because biological systems are inherently complex. models trained on low-quality data sets, as you can imagine, will struggle to generalize.

And also when it comes to large language models specifically, the risk of hallucinations and its non-deterministic nature where it can make up things or not say the same thing each time can be a barrier to adoption in scientific or healthcare settings. Now, approaches like guardrails and other techniques for grounding the AI's response, can help mitigate some of these challenges. But ultimately, especially in healthcare, I think the most effective approach is using a hybrid model where human expertise is kept in the loop by design,

even as newer agentic systems begin to enable more of these autonomous workflows, so to speak. So when developing medical diagnostic instruments for tests, what are some of the practical constraints, be a technical, organizational, or even regulatory that shape how AI can be used? Yeah. First off, I think the main consideration is data governance and privacy, which goes beyond regulations like if our compliance, since many healthcare data sets contain sensitive patient or genome-like information. That affects how data can be accessed and shared or use for model development.

Another key constraint is the system integration. In enterprise environments, models rarely operate in isolation, right? So they often need to integrate with the existing kind of ecosystem. And there are also these deployment and operational considerations where you have to decide whether the AI systems run in the cloud or directly on the instrument. And factors like connectivity, latency, and other associated regulatory requirements, can influence where models can realistically be deployed. And last but not least,

teams need a kind of a post-market surveillance plan, which requires a strong model observability service, where they can monitor the performance of the model and identify any drifts, and also ensure that they are working as intended in real-world settings. So in practice, applying AI in the space means balance, in innovation with these operational and regulatory controls. Now, here comes the hackathons, Janice, so you'll be satisfied.

You've participated in both competitive and non-competitive AI projects, including your win at the BigQuery AI hackathon. How does working within those experimental environments, how does it perform differently than when you're working inside an enterprise, and what does it teach you in terms of what you can bring into your day-to-day work? Yeah, I'm glad you asked. I would say one vessel from hackathals that really carries over is the mindset of rapid experimentation.

You're testing ideas quickly to determine feasibility and you're building a minimum viable product for a demo purpose. And interestingly, my experience with the BigQuery AI competition was that it was more than just rapid prototyping because platforms like Google Cloud provide a very integrated ecosystem. and in our case, we built an AI-powered intelligence pipeline that could look up a patient's multi-oomic data and generate personalized follow-up plans. Particularly, we used what we called as a RAND framework to create supporting literature citations

in order to demonstrate that our pipeline is reliable. So in this way, we got quite close to an enterprise AI workflow in a clinical setting, but that has used that organizations have the infrastructure and resources to support these platforms, right? That said, many research and clinical labs are still cautious about using cloud platforms, especially when dealing with sensitive patient or genomic data. So I think the real difference and the challenge in enterprise environments is the integration with existing systems and the scalability. So in real products, AI has to fit into a larger ecosystem of the software and hardware. And in a lab setting, you often don't have the luxury of a high-end GPU for inferencing.

And this is exactly why we're looking into things like model distillation and quantization, where essentially we're trying to shrink these massive models down. So they're lean and enough to run efficiently on standard, hardware without losing the accuracy that a clinical diagnostic would require. So, Maya, data quality and accessibility are recurring challenges in life sciences. You know, from your perspective, what needs to be in place before AI can meaningfully contribute to product development or diagnostics? Yeah, talked earlier about some of the data quality challenges surrounding AI in healthcare.

What else the first steps to addressing those challenges, in my opinion, is having a strong process for data curation. As labs are collecting experimental data that needs to be well annotated and collected in consistent ways so models can actually learn mobile full patterns. You mentioned accessibility, which is another important piece. In many research organizations, data lives across different instruments, labs and databases. So having that right infrastructure and pike lines to bring those data sets together becomes more critical. I will say that currently one important resource the research community relies on is large open biomedical data sets.

So resources like the Cancer Genome Atlas, which is curated by the NIH, already provides like the identified data sets that researchers can use to develop and validate models. In fact, in my own hackathon work, we also leverage this public genomic data set that's available through Google Cloud BigQuery, which made it much easier to build and test our AI pipeline at scale. Increasingly, cloud data platforms are also enabling organizations to share and even monetize curated data sets in a governed way. So I think that can further accelerate collaboration and innovation on the life science ecosystem. And I think to end it, I'll say, going forward, expanding these open data sets and, you know, enabling federated approaches that allow collaboration without having to share raw data will become really important for AI to start, loophily contribute to these diagnostics and product development. No, I know bioscience is a field where there's incredible collaboration between those in life science disciplines and engineers and data sciences. And I think that other verticals could learn a lot from what has functioned in life sciences as necessity for a very long time.

But one thing that I'm really curious about is how that has evolved when it comes to integrate AI as part of the solution. and what have you learned through that process? Yeah, you're absolutely right to highlight those roles because successful AI systems usually depend on a strong collaboration between them. How I see it done today is that it usually starts with a shared understanding of the problem and clear traceability from your product requirements to these AI features specifications. Domain experts were often scientists or engineers who deeply understand the application, and how customers actually use these instruments,

they help define the scientific question and the real-world use case. And they're also playing a key role in developing some of these high-quality golden data sets, which serve as a ground route for training and validating these models and for fair bridge marking against different algorithms. I do a bit of data science myself, and right now the focus is on developing these models, models and also building robust evaluation pipelines to test the performance of these trained

models and make sure the results are meaningful. And when we say engineers, I'm thinking of both software and IT professionals. They're working together in building this operational infrastructure that's needed to deploy, monitor and maintain these models as part of production workflows. So yeah, these roles are evolving into more specialized teams as AI itself is expanding rapidly. So I can think of teams like data ops and MLOps, just to name of you. So looking ahead, Maya, where do you see the most promising opportunities for AI to impact medical diagnostics and life sciences, research and design over the next few

years? Yeah, I love this question because it reminds me of a line I once heard in a tech talk that really stuck with me. Many of the diagnostic and treatment, treatment capabilities we take for granted today simply didn't exist even for Royals about a century ago. And yet there's still a long way to go, especially in detecting aggressive cancers like Patriotica and Ovidian cancer, where earlier diagnosis can dramatically improve survival rates. I think it's a great reminder of how far we've come, but also how much room that is to grow. So for me, the most exciting area is the integration of multimodal patient-specific data and how to do this at scale. Right now, we're looking at genomics, cell proteomics, as well as clinical

data in silos. And I think AI has this incredible ability to kind of fuse those data together and being able to accelerate key biomarker discovery in ways we simply cannot do manually. So I'll Ultimately, I think it's about earlier intervention and personalized care, especially when it comes to aggressive cancers where timing is everything. So if these systems work as intended, I think they won't just impact patient outcomes. They will fundamentally change how we manage diseases and improves a person's quality of life, which I work is very important. Maya, thank you so much for being on the program.

What you shared was really inspiring to me. and the way that you're leaning into the advanced edge of AI for life sciences is fantastic. Thank you so much for being here and becoming part of the tech arena community. For listeners who want to engage with you, because what you've said is really resonant with them, where would you send them and how would you like them to engage with you? Yeah, thank you, Alison. I hope to be back on the show as well, I think to be your question, as someone who likes to learn by doing my advice,

would be to focus on platforms that let you build practical skills through like small hands-on projects. Working with like real data sets and experimenting with these models is one of the best ways I think to understand how these AI systems actually work. And today there are many tools that make this easier. For example, like cloud platforms like Google Cloud often provide free trial or commercial credits that allow you to experiment. with services such as Vertex AI, and you can build end-to-end machine learning workflows within that system.

I often go to a platform like Kaggle because it's helpful where they offer a simple environment with a limited free GPU, of course, and that's great for testing models without needing your own GPU infrastructure. So I also find it helpful to follow a few AI educators who share practical tutorials, and insights, again, since the field is evolving so quickly.

I'd say in short, the key is to keep experimenting and just learning through practical projects. Awesome. Thank you so much for being on the platform today. And Janice, that wraps another episode of Data Insights. Thanks so much for being here with me as we explore the next generation of AI. Thanks for joining Tech Arena. Subscribe and engage at our website, Techarena.com. All content is copyright by tuckering.