Transforming Life Sciences: AI, Vibe Coding, and Drug Development Acceleration

Transcript

Shane Hastie: Good day folks. This is Shane Hastie for the InfoQ Engineering Culture podcast. Today I’m sitting down with Satish Kothapalli. Satish, did I get that right?

Satish Kothapalli: Awesome. I think you are one of those first ones who got it right.

Shane Hastie: Well, thank you very much for taking the time to talk to us today. My normal starting point for these conversations is who’s Satish?

Introductions [01:22]

Satish Kothapalli: So okay, who’s Satish? Let me start, this being a technical staff, I’ve been in the IT industry for more than 25 years now. Like most of the IT folks, I started as a co-developer. I think it was C, C++ days, right? It takes me back to 25 years, C, C++ days, right? So I was a developer, followed the same regular pattern, senior developer, technical lead, an architect where I was mostly focusing on the content side of it. I moved into the enterprise content management. Then I became an enterprise architect specific to the content space. During this process I was working with a lot of different domain companies, which includes manufacturing. But licenses caught my interest and I glued to licenses for more than 20 years now.

So I worked in a license company for more than 10 years where I was doing the enterprise architecture work. Then I moved into the automation. I was heading the content and the automation. Then comes the gen AI, right? So automation content, gen AI can’t go anywhere other than mine, right. So I was hitting the enterprise content automation, the gen AI platform services for the pharma company, and then I moved out of it recently, it’s just been five months I moved into Altimetrik. I’m working as the CTO for the life sciences business unit there. I’m basically spearheading the adoption of the latest cutting-edge technologies, specifically focused on delivering transformative solutions to the life sciences.

That’s in short what I’ve been doing in the last 25 years. I’m very happy to be here for one reason. So while I was a coder, I always used to think about this. Why am I not able to teach computer my language? Why am I following it and putting all the hard work. So we are at that stage where we are talking to the language, language of the computer. Computer understands my language. So very happy to be here part of this session, especially on the vibe coding. Thanks for inviting me here.

Shane Hastie: I’d like to dig first into life sciences. What is special about building software in the life sciences space?

The Role of IT in Life Sciences [03:47]

Satish Kothapalli: This was very interesting one. So even my kid used to ask me, “What do you do, Dad? You’re part of an IT in a life sciences and the life sciences is all about making the medicines, pills, drugs for some diseases. What do you guys do?” I really struggled to answer that for quite some time, but now I’m proud to call it out that IT professionals do a lot of things in the life sciences space. For example, we’re talking about close to 12 to 15 or even beyond to get the pill out into the market. It’s a compliant, we have to do this, that, prove it for more than 15 years before the pill comes out for all the good reasons. So we are not questioning any of those.

Accelerating Drug Development with AI [04:33]

But that process is getting automated. We have been trying to automate, we were able to reduce by weeks, months. But with the AI, especially the gen AI, we are able to cut that down by years. And again, we are talking about one day saving millions of lives here and we are talking about technology cutting down by years, coming from identifying the right molecule to solve a specific disease or let’s say identifying the right protein and identifying the target. We are at a stage where we are able to cut down from two, two and a half or even four years to a few months going forward.

Then we talk about tons of documentation. In fact, the life sciences, 25% of the total clinical trials, which is a couple of billion dollars, usually 25% of goes into the documentation, review, creation, all this stuff. Thinking about the gen AI, cutting it down by 70, 80% soon, we are still talking about 40 at this point of time, but it’s going to improve. What does it mean it’s cutting down the timelines of the 15 or 12 years, under 10 years sometimes soon. Again, I’m proud to be part of this life sciences, especially in the IT space because we are doing the magic now.

One quick example. So previously we used to have the scientists, the PhD folks identifying the structure of one protein in the complete PhD work, and Isomorphic Labs did that in couple of weeks for millions of proteins, their structures. That’s like millions of scientists coming together and working for years is done over a weekend.

Shane Hastie: Massive acceleration, improving lives quicker through that time to market. The skeptic in me wants to say, will the AI be as cautious as the human being?

Human Oversight in AI-Driven Processes [06:41]

Satish Kothapalli: Again, we are not saying that… Let’s say when I say the created document or review in the clinical trial space, it’s giving me a first draft of the document, which usually might take around months, but this is there around a couple of minutes. So human in the loop is definitely in there for all of the activities, what we conduct in the life sciences space. We are talking about productivity here. We are enabling the user to be more productive, but we are not replacing any of them. So the process is shrunk because of the productivity that we are enabling at each and every step. So don’t worry, we are good. We are in the safe hands of the scientists. It’s just the productivity that we are enabling there.

Shane Hastie: Cool. So leaning into the topic that we connected about, vibe coding. It seems to be a buzzword. Everyone’s doing it, but are we actually getting benefits? How’s it going and what’s happening?

Understanding Vibe Coding [07:41]

Satish Kothapalli: Okay, let me start with what’s vibe coding for some of the folks. As I said, from the day one, it was especially C. C was more complicated, complex, and it was tough for me to learn. Then when I saw C++, I was like, wow, it’s way better than C. Then it was Java. I was like, wow, it’s even better. But always I had this, why am I learning the language of a computer? Why it’s not learning mine? And I feel at this point of a time, we are at the stage where through vibe coding, I’ll go back to the term vibe coding. Again, I really love that term, but I have alternative buzzword for that. The vibe coding is actually being the bridge for me as a coder and the computer. It’s enabling my language and converting it to what computer understands.

So previously I used to take a couple of months to build a simple application. Let’s say I still remember doing this, the school management system for a school which took me around four to five months and I got good money. I’m not saying that it was bad, but I was able to charge them for five or six months of effort. But I tried that with the vibe coding and it built way better than what I did back then, taking six months in few minutes. It’s able to improve the productivity of a developer. Again, vibe coding is very commonly used at this point of a time. So we were talking about the virtual assistants to start with. When we heard the gen AI, where we were wowed saying that, okay, it can understand me and it can answer me. Then we kept on expanding it and we said, okay, there is an agent possibility where it was started acting based on my question of the request.

So then we said agentic AI, the next level, where we said there are three or four agents working together and getting a goal it choose for me. It’s exactly the same, but it was very specific to a developer. So that’s agentic AI model, which is doing code for me or for a developer is what it is called as vibe coding today. So coming back to the vibe coding and the term, I really love this guy Andrej. I keep watching his videos a lot, great videos, but while we talk about the other agents as copilots, coworkers, and I didn’t understand at the time why this term vibe coding for it, but I would still say these are going to be the coworkers. I don’t want to look at it as a tool or a terminology or a technology, other word. They are the coworkers for any developer.

So you’re going to get to have four or five coworkers and you’re going to orchestrate on what they want to do for you. What we see here is most of the companies are still in the experimental mode and we are seeing a few companies already completing the experimentation mode and moving towards the integration. We see a lot of articles, buzzwords here and there. It looks like we are buying the speed from the future. You get what you want to do quickly now, but you might end up into technical depth where you have to spend a lot of time cleaning it up when you get into the next model.

That was true, that was very true in the beginning days, but things have improved over the period of time. Last year when we had this word, it might help me up to 30%, 40%, but the days have changed and we all see how quick all these things are changing and almost everyone are betting on this vibe coding tools. So even recently AWS, Google, and we have all the top LLMs being part of this, enabling that speed and other stuff.

So the important thing, what I have seen is we are still stuck to that age-old SDLC life cycle where we want to keep all this stages of the SDLC. Okay, you know what? There is a requirement phase, there is a design space and then we do an architecture, then do the coding testing. What we are talking about today is once we have an idea, we definitely need to work with the business product owners trying to get more details. But even AI is helping that fine-tuning and providing some new features which we are not even thinking and able to convert into the detailed designs. With the help of the current architecture team, we can just call it out what is the architecture patterns we follow for certain kind of an applications and just pass it on. And we get to see at least the 70 to 80% of the code, which is a working prototype with all the functionality baked into it. So now the developers can just go ahead and fine-tune it to make it production ready.

There’s always this question, is it production ready? No, it won’t even be production ready in the next two years. We definitely need a developer looking into it, fixing some of the issues and making production ready and scalable. That’s still in there. But what changed is if a business analyst or the product owner have a clear understanding on what they’re trying to build, they are given this magic wand of the vibe coding where they can build the application because they know in and out. Instead of writing the documents, which we use to do it for weeks, months, they have a solution in place which is working more or less to what they were envisioning and hand it over to the developers to fine-tune it.

So the developers now are becoming the orchestrators, I would say this is where we are heading towards. They’re mostly becoming the orchestrators and a validator at the end once the coworkers build the applications for them, right? So that’s where we are, definitely some guardrails, cautions to be in place even before we jump in and start building applications using the vibe coding. There’s a lot of white papers around this, the videos around this on what those guardrails are. And we are talking about the security guardrails, making sure that we are not exposing some wrong code, the prompt injections into the code built by this. But again, it’s not providing for sure a production ready that we can do a click of deployment into the production. We still need a developer who is able to orchestrate and validate the output and fine-tune for the production.

Shane Hastie: There’s a lot in there to unpack, but let’s start with that shift to the orchestrator role. If I defined myself as an author and as most software engineers think of themselves as creative authors of this work, there’s a philosophical shift when I become an orchestrator, isn’t there?

Coder to Orchestrator [15:06]

Satish Kothapalli: It is and that’s exactly the problem we are dealing with with most of the companies today, especially the developer. Me being a developer myself, I can feel, but it’s a technology, we need to adapt. All right? So I kept on hearing this, especially the gen AI space, right? It’s not the gen AI who is going to replace you, but it’s the folks who understands and use the gen AI are going to replace you. It’s exactly the same thing. We are not talking about replacing a developer here. You get to do more better work in a short time.

So basically I would say you are being promoted to a technical lead now because you have three or four coworkers who are going to work for you. You are defining the design, you are providing the architecture, how to deal with it, but at the same time you’re managing these coworkers or the kids saying that okay, this is what you have done. Let me take a look to see where did you fail and I give you the feedback but I correct and I validate and move it to the production. So that change, that adoption is something which is slow at this point of a time, but will definitely happen soon.

Shane Hastie: So I can see that and an experienced developer can step into that. They’re used to reviewing code, they’re used to providing feedback. So if I’m a team lead, if I am in that tech lead space, that’s a big part of my job, but what do I do with a novice who hasn’t learned how to do that yet?

Learning to Learn [16:40]

Satish Kothapalli: Now you’re right. There is a definitely lot of upscaling that needs to happen while you’re focusing just on the developer at this point of a time. There are a lot of roles that will be changing right now with this complete SDLC life cycle might be shaken because it’s providing the ability even for the business analysts where they have to upskill themselves to understand some of the architecture and the designs what we as the developers are talking today. So they need to know because they have to start interacting with these vibe coding tools when they are doing it, while there are the templates designed and given by the developers to them, they still have to talk to them, get the things done right. So they have to upskill slowly into that space. That’s definitely the step one for the business analyst, but coming to the developers, you’re right, especially the folks who recently came out of the college who have been trained only on the development, but they are not great in the area where they are.

They’re just coming out of the college. They’re learning. They slowly have to start making a shift towards this because you know that the space is changing and at the same time the gen AI is offering all the right tools for these folks to make a shift. So again, I’m not seeing someone have to manually do the complete validation to themselves. Again, the agentic AI comes into the picture, trying to help figure it out and you can just ask it, “Okay, here is my application, can you take a look to see is it following to the best practices of what is already given by me or the architecture standards of not being a monolithic but a cloud-based services?” And it automatically detects and identifies and provides that, right?

We are shifting. The earlier we learn and upskill, especially these folks who are in this mid-range is good. I take sessions to these guys for the weekends as I was talking to you offline, this is this piece. We definitely have to hand-held them. I do my bit of it. I take the weekend sessions to them trying to explain because they’re still in that mode of denial. No, no, no, no, because we saw an article saying that it’s not going to happen for the next five years while some of the gen AI leaders call it out, it’s already in there. So there’s a lot of different messages out there. So how can we filter it out and help these developers in that segment of zero to four or five years of experience? There’s no other way. We have to up-skill them as quickly as possible.

Shane Hastie: So use the tools to validate the output of the tools. That feels almost like it’s cheating.

Satish Kothapalli: It’s the technology. So you still have this idea on how to build a solution and it’s enabling the next-level innovation. We were spending 70, 80% of our project work just doing this, knowing we know what to do, but just to do that. So we kill 70% of the overall time, but think about this, the technology is going to enable these developers to build total different kind of solutions. It’s going to take to the different level. For example, when we have the quantum computing coming in, we need developers again. It can’t do it because it learned from what was already existing. Any new architecture is something that comes out of the brains of these developers of the architects. So they get to do much better work, instead of doing this mundane 70% of the work, now they get to focus on better real-world challenges to deliver the solutions quickly.

Shane Hastie: So that’s that abstraction up instead of the human being learning the computer language, the computer has learned the human language. You made the point that other roles in the software development lifecycle will be doing some of that. What’s happening to the software development lifecycle and what’s happening to the nature of these roles?

End-to-End AI Augmentation in SDLC [20:59]

Satish Kothapalli: Good one, it’s still relatively new I would say because we didn’t want to adopt it and we have been dragging this for quite some time. But as I said, some of the companies are already into it because we see the customers asking us, “Do you have a vibe coding community of practice or the COE where you can give me the productivity and the speed that what we are looking for?” So for now, what we are looking at is a business analyst, and again, even the role of the business analyst is expanding a lot. Like I said, while it’s moving towards understanding the technology, but it’s emphasizing and forcing this business analyst to be strong in the domain space as well. Understanding the right problem and understanding how it can be achieved or implement the solution is going to be the key here. So the domain expertise and the problem-solving thinking is getting critical.

So business analysts sitting with the business who are actually doing life sciences, identifying molecules and other stuff, the business analyst understands with the deep knowledge and expertise he or she has and converting that into the requirements. And again, gen AI is actually helping us out there as well. But once we have the business requirements outlined, at a very high level, breaking down into the stories of the tasks, it’s something the gen AI is able to offer as of today. It might not be a hundred percent accurate, but as I said, it’s giving the first draft of all these documents with this specific set of the stories with the architect and the developers team coming in saying that this is our technology stack in our company and this is our architecture patterns.

Just asking the gen AI saying that, “Okay, here are the stories, I need to implement a solution and here are the guardrails for you from the architecture and the design standpoint and the security controls. Can you come up with a design document and the architecture document?” So again, it’s a first draft. I’m not saying it’s replacing anything. It’s a first draft as good as maybe 60, 70, 80%. It’ll definitely keep increasing, but we will not be able to skip that step and say we automate. We are just augmenting that at this point of time.

So once we have all these details of the architecture, the design, handing it over to the vibe coding tool that’s generating that solution close to 60, 70 is what I would say at this point of a time to what we have defined in the design spec, the architecture and the storyboards, and the developer comes in and he says, “Okay, I can fine-tune this. Let me fine-tune some of the stuff it missed on this. And recently what I’ve seen is some of this vibe coding tools have integrated with the browser MCP.

What does that mean? Basically while it’s creating all these documents, while it created the storyboard, it even creating the test cases for us and the test scripts, and when I asked the vibe coding tool, it’s executing this functional script, opening the browser and detecting if there are any functional errors or the technical errors. Because it’s able to access it, it’s fixing the code on its own. It’s augmenting each and every step, each and every role. Let’s be on the safe side, let’s say 50, 60%, but it’s augmenting and cutting down the efforts and improving the speed and the capability. We are seeing how we can change some of the roles, titles and other stuff. It’s the business analysts, the orchestrator. I don’t know what we are going to call the QE stuff because this is able to do most of the stuff. It can do the functional testing, it’s automatically fixing it. That’s a good thing.

Now, once we dump this into the production, we are saying, okay, let’s have a log collector collecting all this information and we have an AIOPs sitting on top of it, which is basically looking at the logs and says, “Okay, you know what? Looks like this is heading towards some problem because it’s consuming a lot of memory or the CPU usage, whatever it is, and it’s even suggesting what needs to be fixed where.” And if something fails in the live system, this AIOPs is actually detecting this functionality is failing for so-and-so reasons, go ahead and fix it. Previously I used to take weeks when I have the multiple systems talking with the API connections. Now I just collect the logs and paste it in here. It’s says, “Okay, now I see what the problem is, let me fix it for you.” It’s going crazy. Sorry to say that, but yeah.

Shane Hastie: These are huge shifts. You’re making the point it’s 60 to 70%. There’s still the need for that human oversight, the human critical thinking and it’s a different world, isn’t it?

Productivity Gains and Workforce Implications [26:16]

Satish Kothapalli: It is, it is. Now while I keep myself, you’ll have something else to do, the only one question that keeps coming to me is if I am improving the productivity of the user by 70%. Let’s say if the work is done by 10 folks, the same work can be done by three folks going forward. So whatever we see, I don’t want to sugarcoat any of this stuff. The thing is we are assuming, okay, that 70% productivity are the cost savings, whatever we want to call it, would be invested into the next level where we would be thinking about bigger problems and the bigger solutions where this seven members can go in. That’s something I don’t know at this point of time. I’m hoping that’s going to happen and we are going to find a lot of good things going forward with the productivity what we are gaining here.

That’s not something I can comment at this point of a time, but I’m very cautious and I keep my friends, all the students who comes to some of my sessions cautious about that. Make sure whatever you’re learning, you are not scratching on the surface of any of this, but you get into the deep depth of it because this vibe coding of the gen AI is not able to do anything beyond 70, 80 or maybe 90, right? But the rest, 10% or the 20% is forever there and you’re going to get that expertise. But while we are doing that, even the orchestration is very important. And believe me, it can throw you 100 different solutions for the same request unless you do the right prompting, right, the guardrails, the architecture, scaffolding and other stuff. So even that, we need good resources there. So it might be not eliminating the roles altogether, but I see a change a little bit might be ending towards the future solutions. So I’m still hopeful.

Shane Hastie: So training those agents, giving them the guardrails, giving them the design patterns, the way we do things here, how do we do that?

Training AI Agents with Guardrails [28:39]

Satish Kothapalli: It’s basically same something which we do with the kids. So kids start coding and what is some of this LLMs? It is all this repositories, which we kept open, public and all this were trained on. GitHubs, all the free GitHubs, or even some of this websites where everyone started posting, okay, here is the error what I get, and there’s a solution, marked as a solution. All these are good information that went into this LLM models and trained. So it’s writing that code. So now when I say, you know what? I’m on AWS, or let’s say I’m on Azure, not on the GCP or AWS, the context which I’m passing it that I’m on Azure services, not on AWS, is clearly calling it out. I need to start looking into the Azure services in order to bring this in and they’re plugging those services and that kind of a code into it.

And is it always doing a hundred percent accurate? It might not. This is where the developer still comes into the picture, right? For example, I’ll give you one thing, right? So we keep hearing this LLM so-and-so model and now solve something in 10 minutes where PhD folks took two months or something. So as the same LLM, I tried doing that, okay, I have a problem here. I ordered shoes from Amazon, I got a left shoe instead of right and a right shoe instead of left. What should I do now? It says, “Oh, sorry to hear that. Here is the Amazon contact center information. Usually they take one day to come and pick it up,” and all this. It’s the training. If they are not trained, for example, this is where I was saying if something new comes in and then some new architecture comes in, it doesn’t have a clue. It has to come from the architects, it has to come from the developers, the designers. Once it learns over the period of time it can bake into the solutions.

So it’s still dumb. While it fakes it out that I’m too brilliant, it’s just following a pattern of identifying the next token. It’s exactly the same. Because I said I have a problem and I said Amazon, it has this picking of the words identifying what is the right word I need to use in my context to find the next word. It saw Amazon. It said, I have a problem, I want to talk to the customers. It said, okay, here is the problem, left and right are not the words. These are the words. Then it started thinking, okay, when I put all these words together, the next word is the reach out to the contact center or the phrase. So this is exactly where we need all this human intelligent brains to come and solve it.

Shane Hastie: Satish, a lot of great thoughts and food for thought in here. If people want to continue the conversation, where can they find you?

Satish Kothapalli: I’m very active on LinkedIn, keep posting some of these articles and I even do some of the sessions for the weekends, like I said, for the kids coming out of the college with my friends, and we are going to a place in India, Vijayawada, where we want to enable the students who are in third year, fourth year enable with some of these things how to shape the learnings for them.

Shane Hastie: Wonderful. Thank you so much for taking the time to talk to us today.

Satish Kothapalli: Thank you so much and it’s a pleasure being here.

Mentioned: