Navigating AI Implementation in Pharma

Daniel Levine (00:00) The Life Sciences DNA podcast is sponsored by Agilisium Labs, a collaborative space where Agilisium works with its clients to co-develop and incubate POCs, products, and solutions. To learn how Agilisium Labs can use the power of its generative AI for life sciences analytics, visit them at labs.agilisium.com. We've got Vinod Das on the show today. Who is Vinod?  

Nagaraja Srivatsan: Danny, Vinod is an expert on AI and data science with over 20 years of experience in the pharma R &D space. He's currently associate director at Bayer Pharmaceuticals, where he's responsible for the enablement of AI in their group. He's keen on driving the integration of generative AI and large language models to accelerate drug discovery and clinical development. I'm really excited to have him on the show.

Daniel Levine: And what are you hoping to hear from Vinod today?  

Nagaraja Srivatsan: Vinod is a very practical and pragmatic implementer of AI. And so what I'm really looking forward to learning is how he applied that framework on selecting the right gen AI use cases, and then how he built the capability within his organization to implement them, and finally, how he's gone about evaluating its impact in his organization.

Daniel Levine: Before we begin, I want to remind our audience they can stay up on the latest episodes of Life Sciences DNA by hitting the subscribe button. If you enjoy this content, be sure to hit the like button and let us know your thoughts in the comments section. With that, let's welcome Vinod to the show.

Nagaraja Srivatsan: Vinod, welcome to my podcast. It's really exciting. Today, I want to explore some of the AI use cases you've used in your organization. It would be wonderful for you to describe the before and after. What was the situation before AI was implemented? And then what was the journey for AI? And then what was the situation after? Thank you so much, Vinod. Over to you.

Vinod Das (02:04) Yeah, thanks, Srivatsan, for having me here. And it's a splendid opportunity for me as well to see and share the AI journey to reflect on what we have done from the Chat GPT moment in 2022 to where we are right now. So we went across that short-term, mid-term, long-term strategy. And in short-term, we were pivoting around, OK, what is the best fit for us? For example, protocol authoring. OK, we work with business functions to see hey, we have a lot of protocols which get amended. So amendment cost is in millions. So why not reflect and see the tangible outcome of it, right? So we strategized that and we partnered with external to make sure that we are effective and we are accelerating the deployment. So that is super helpful. And we had it in Sprint, say Sprint 1 runs for four weeks - like sprint towards specific peaks, so that we see some tangible outcomes. So that's where we were able to show and tell the user personas, user stories, and reflect on what is really required for a user journey to have a proper protocol graph prepared.

Nagaraja Srivatsan (03:18) Vinod, that was excellent. Tell me a little bit about the outcomes. What was the outcome measures before and then what was the outcome measures after you implemented? You talked about measuring key outcomes. So why don't you touch upon what were those outcomes and how did you measure it?

Vinod Das (03:34) Yeah, that's good question and also intriguing as well because when you really interview the medical writers or clinical study managers, you get some person, this can be tweaked or this section can be tweaked. And when you really reflect it on the draft, the LLMs hallucinate. So how much human intervention is required. So we gauged it in that perspective. And when you really see the tangible outcomes perspective, it reduced the productivity aspect by around 25%. We cannot minisculely say, OK, it reflected 90%. No, not like that. 25 to 30 was the acceleration which it gave.

Nagaraja Srivatsan (04:25) No, that's a good narrative saying that, hey, first bring the users on board and they are giving you qualitative and quantitative measures on how good the output is. And the second you said is that as the users started to give feedback, it was helping them be productive, but in that 25 to 30 % range. Did you roll it out to a lot of people or was this a pilot for a few medical writers? What was your strategy? Did you do a pilot with start small and then scale?

Vinod Das (04:56) When we started, we started off as a proof of concept, then moved to proof of value, then we deployed it across the work streams. That is the global application, which is available for production use. So it can be utilized across bio for utilizing it and seeing the value.

Nagaraja Srivatsan (05:15) So tell me a little bit more of the nuance and mechanics of that - was in that proof of concept, did it take four weeks, 12 weeks, 16 weeks? What was the timeframe? And then what kind of tools did you use? Did you use out of the box LLMs or did you have to use others? Walk me through the journey of how did you build this?

Vinod Das (05:37) The build phase is very exciting because we had a agile workstream. So we had end-to-end within six months, right from, say, developing the backend and the frontend, where users can invoke a prompt and get some tangible output. And later on, always there is an incremental progression on the UX based on the feedback, where we had a prompt templates developed. So that's how we went in iterations. And when you see the usage, yes, they do have good metrics on the usage side of things. Only thing is it is constrained to specific, let's say, transcelerate template, for example. So that is used across the board in different organizations. So we have used it in a backdrop templates to understand the protocol nature and reflect on it. So when you say model, we used OpenAI Chat GPT to start with because that is more intuitive in 2023 and more user reflective.

Nagaraja Srivatsan (06:50) You brought up some very good points, Vinod. One was this whole notion of prompt, prompt library, all of that. It seems to be that prompting is the new way. Tell me about how, you know, we're all from the IT background, we have coding standards and others. How are you doing that with prompts and that notion of prompt and prompt libraries?

Vinod Das (07:12) That is also very important and very useful because the business context or the ground truth is derived out of it, right, for the model to learn. Say for example, Srivatsan is talking to Vinod Das because Vinod has some personality and you're putting some tailoring questions out of it. So similarly for the model also the same reflects out of it, right? So we were putting across, okay, for this protocol authoring, what is the informed consent information or what is the primary endpoint? What is the secondary endpoint? So we were able to do like one-shot prompting. There is a chain of thought prompting, there is atom prompting, now there is a chain of draft promptings. So there is, it's accelerating. Prompting is important, but business context also lies closely with the prompting aspect. So we married everything to have okay these are the golden prompts which can be readily used by the medical writers for generating a draft.

Nagaraja Srivatsan (08:10)

It was excellent the way you talked about how you created that prompt culture with both the prompt techniques of zero-shot, single-shot, multi-shot, but also how you train them in the business context. Tell me, how did you approach it? Were there standard courses, standard training material? How did you build that capability within the organization?

Vinod Das (08:33) So Sri, is also a very important question for other organizations also to piggyback or to use these learnings. So we had, or it's not me, the team, there is a learning and development team who build this prompt journey of prompting techniques, exercise which starts from square one, the basic level of prompting to advanced prompting and we internally embark and get certified. It's through an internal platform called MyGenesis . So that team worked rigorously to build this curriculum. So this helps for AI literacy perspective as well as understanding what each model does. For example, we have Gemini, have GPT, GPTs, and we have Anthropic suite. So we have multiple variants, right? So each model, how they behave, behavioral pattern of it, et cetera, et cetera, was tailored for us in the prompting.

Nagaraja Srivatsan (09:29) Vinod, that's an interesting point you're bringing in of each of these different models. Do you have a standard prompt across each of these models or do you do differently? And have you used the same prompt to different models and are they giving you similar results or different results?

Vinod Das (09:47) That is also a very important question because when you see the dataset behind the scene is different in which the models are trained on. Say for example GPT-4.0 versus GPT-4.0 Vision versus say Anthropic 3.7, Cloud 3.7. each dataset is different. the underlying, say captions, or let's say the syntax of the prompt will be the same. Only thing is we tweak it based on the model behavior. Say I want to generate the primary endpoint for this particular study. For example, I invoke that and the model output is a little bit sketchy or doesn't give me a proper output. So in that case, we tweak it to say, can you be an expert in clinical protocol authorizing and do this task. So we do the star format, S-T-A-R, situation, task, action, and the result out of it. So we follow the STAR format, so it's more reflective from a model side of things.

Nagaraja Srivatsan (10:56) You that you bring up a very good point because most people use the STAR framework for interviewing and others where you describe the situation, the task activities and results. You're bringing an interesting twist that that format could be a good scaffolding to build good prompts because basically you're telling the prompt, this is the business context, which is a situation. These are the tasks you have to perform. These are different activities which you have to do, and then these are the results you need to get. And you can then bring in your ZeRO-shard, ZeRO-1-shard learning in each of those components. I think that's a very wonderful key takeaway on how you build your prompt and prompt capabilities. Tell me a little bit on this L &D. Was that an internal course or, as you talk to other people, that are external courses which you would recommend or you've seen? From that people can upskill themselves in this particular new technique.

Vinod Das (11:55) Yeah, so this is learning, it's an internal system where our MyGenesis system has developed and they gave it to us with a certification behind the scene once they complete all the journey. So it's co-created by a constant feedback within the employee, I would say, cohorts, right? So that is one. And when you want to see in the marketplace, like when I started, I went to Udemy. Udemy has a very good cost structure on prompting and the prompting techniques and what sort of prompting is good prompting for a model. So those are all small tips and tricks. And apart from that, I would strongly recommend you to be a model, like I would say, into tests. Say for example, each model behaves on its own. So have some template structure, document it, and have it as your cheat sheet so you can immediately do some tweaks and tricks to suit the model needs.

Nagaraja Srivatsan (12:59) And Vinod, I know we're exploring prompting because that was a very critical part of your protocol authoring journey. How did you ensure version control of these different prompts? Was that a little bit of, you know, are you doing it in Word? Did you have a kind of a platform to do this?

Vinod Das (13:17) That's where the ground-proof prompts, which is behind the scene, which cultivate the model capabilities. So each medical writers or study managers can put their prompt also. They can have their prompt. Similarly, with the MSCo pilot, for example, if you have your prompt, you can save your prompt and invoke it on need basis so that doesn't impact the back end of the model, but you are interacting with the model for that part. So from a versioning perspective, we have the ground root of prompts, which sits behind. And on the front end, there is a user modeling interaction, which is basically in the front. And you can store those golden prompts, I would call it.

Nagaraja Srivatsan (14:08) That's excellent. I think exploring your prompt journey, which is so important, was wonderful. Vinod, I know you've talked a lot in the marketplace about agentic architecture and your implementation of agentic architecture within your organization. Walk us through how do you start thinking about an agentic architecture platform or set of use cases for clinical development and just walk me through your thinking.

Vinod Das (14:37) You're pricking my brain, because agentic architecture is evolving. You recently heard about model context protocol, which is the USB for LLMs. So there is always an acceleration in the agentic  side of things. And LLM architecture, you think about it, it always revolves around the data and how normalized is your data and the cognitive aspect of your data, say my knowledge extraction and knowledge retrieval, how can I blend those both? You can extract the data and keep it, but when you want to retrieve it, how many tokens are consumed? What is the cost which is borne on the consumer side as well as from the model provider side? So all that factors from when you do the architecture design. And also this reflects on also the scalability and re-usability components, right, from architecture side of things. So there's a cognition layer, there is a data layer, and there is a, I would say, a scalable, reusable layer. So everything revolves around some reward function to the model. Hey, you did a good job, take this reward, right? So the model understands more, these guys are praising me, why not I do more? That's where you do the fine-tuning, so that the model has good movements within itself. It's like a human. It also reflects what you do.

Nagaraja Srivatsan (16:10) Yeah. you know, you're tasked to do multiple different use cases, how do you evaluate what is the right use case for generative AI? Do you have a framework? And if you do, how do you apply that framework to select the best use cases?

Vinod Das (16:29) This is more reflecting on the tool selection side of things. So you can observe the full quadrant approach, which means my left quadrant is more into problem urgency. What is the real core narrative for my problem? And what is the market situation for the problem? And how can I value it? That is where the problem urgency is. And also, that also has the ideation, whether it's a unique one or can I buy versus build aspect also comes into play a little bit. And the next thing is on the solution efficacy. If I give the solution, how does the user community react, whether it is segmented to certain populations. So solution efficacy also plays a vital part. And the below quadrant is more into, OK, I have my solution idea and I want to implement it. What is my implementation complexity? Whether it's a easy low-hanging fruit where you can just turn on the button and it fires up, or it has a complexity of additional external plug-ins or external libraries which you need to acquire or procure or configure and then build it on top of it, or your data is super siloed that you need to understand the business logic more. There's a lot of implementation complexity along with it. And finally, the most important, the cash cow, which always says some maintenance burden. What is my overhead to it once I implement it? And I want to scale it. Whether there is a regulatory aspect when I deploy it in, say, for example, China or Asia countries. Or is there a policy or regulation which does not allow me to implement it. So there a lot of maintenance burden to it. So these are the four pillars or four quadrants I would say.

Nagaraja Srivatsan (18:18)

And those are very, very, very practical and pragmatic quadrants, You're looking at the business problem, the solutioning, you're looking at the implementability complexity of the platform, or in the maintenance complexity of the platform. So if you put this score for your protocol authoring, walk me through each of the quadrants and how did you evaluate each of them? Where did that fit in? How did you select this as a very good key target?

Vinod Das (18:50) When I was evaluating this, the protocol amendment was reflecting around the score of 0.72. So I need to reflect on protocol authoring also. I need to marry both so that the score is more, like for example, I crossed more than one, so one scale. So what that means is there is a small rubric guide which will say, if you go with this approach, your return on investment will be good enough for consumption. So that's where each quadrant was evaluated. And there is a rubric which plays behind the scene.

Nagaraja Srivatsan (19:31) No, I think it's a very fair business problem. We all know that protocol authoring done right, the digital protocol is the holy grail for what is being done, and medical writers and medical authors take a lot of time, and it's also a change management effort. Before I go into the change management effect, which is a big one, I want to talk to you about that. I want to go into the two parts, the implementation complexity. So walk me through how the implementation complexity was for this particular platform. And also talk about maintainability because you know you're maintaining it as you scale number of users and stuff. Give me a little bit of a parameter on what you had said financial cost of these implementability and maintainability. Just give us some ballpark on both the implementation complexity and the maintenance complexity were evaluated.

Vinod Das (20:19) When you really see the numbers, it went to half a million when we started. Then the maintenance burden, another close to 300k. And these are all early times, 2023. So when you see as we develop the new functionality, there were changes in the marketplace, for example, Instructure.io or for parsing the content. For example, it has tables, lists, and figures. So during that time, it was not able to understand the tables, lists, and figures in a good way. So that is the implementation complexity. So we will go with only text embedding model. So those are all some, say, certain areas where we need to adjust and see, OK, the market is evolving. But we have to definitely go and progress on it. This is the implementation complexity side of things. And when you see the maintenance burden right now, it's more into the advancements. We didn't have the agentic framework, as you rightly mentioned, on the architecture. So now we need to do the reverse engineering or re-engineering to a little bit change a lot of things so that it suits the agentic frameworks. For example, the orchestrator, which can understand the persona for study manager or a medical writer, then it runs to supervisors and after it initially drafts, then whether the study design and my plan is optimized, yes. Then the human in the loop checks it and gives it to the next layer. So those things are not there, right? So this is where the maintenance burden is, right? We need to go with the market. Yeah.

Nagaraja Srivatsan (22:02) So you're bringing up very fair points, especially the human and the loop. So I want to go back to that previous question. In any of these AI projects, there's change management involved, because people are set in using it in one way, and now they have to use it in a different way. Talk to me about that change management process. What was it before, and how did you make people get comfortable to this new way of doing it, and then get them to the...the solution to adopt and use the solution.

Vinod Das (22:30) Yeah, so change management is culture oriented, right? I want to use it, right? And I want to see what's next for me when I use it or whether it is going to affect my job, right? Et cetera. There's a lot of factors to it and also the mindset of it, right? So like, there is a before aspect, before this particular protocol or thing was there, okay? There was a lot of manual work between different members of different colleagues. And after this, it's more into, OK, the technology is automating it. Can I have a shared, say, prepared? And then collaboratively work for the outcomes. So it was in that angle. And when you see the adoption, adoption is there. Only thing is the training which is required. So they prepared some quick reference guides with the help of the developers. So we understood easily what areas to click on and how to generate it, and what are the prompt library which can be already used and what you guys can already create. So we had a quick reference, the Tides QRGs, to show and tell, and we had many live sessions to educate the...clinicians. Let me put it that…

Nagaraja Srivatsan (23:51) No, that's very good points you're bringing in. So first is the before, but then after. One is training, show and tell, giving them assistance in terms of these prompt libraries, and really enabling them to do the job in a much more efficient way, but also giving them empowerment so that they can tweak and do it by themselves. I think you hit upon a very good part where you're not being prescriptive, but you're giving them a good start, which then they can continue to improve and scale and share. So Vinod, I know that protocol authoring is one of many such very innovative projects you're doing. Is there another one you want to showcase which talks about the journey of before and after?

Vinod Das (24:35) Yeah, probably I'll bring one interesting topic on secret data or sensitive data. Majority of the pharma companies are trying to achieve success factor around it. So probably that is one thing which I thought, okay, let me bring it up. We haven't achieved the full end goal, but we are marching towards it. So probably that is something which I can discuss with you further.

Nagaraja Srivatsan (25:01) Yeah, why didn't you describe what that is? What is the business case using your four quadrants? What was the business? What was kind of what you were going through? Maybe that's a very good journey for you to share with us.

Vinod Das (25:14) The problem urgency is we have millions of documents, It's just structured, semi-structured, unstructured, right? So this is sitting in a secret repository and sensitive in its proprietary nature, right? So you are tasked with querying results, say for example, sorbital, right? the liquid alcohol, right? And you get around 200 plus documents when they query it. Sorbital, say, give me the sorbital solubility for 1.5 % or something like that. So it will give around 200 plus. So this is the problem which we had. So the scientists have to go one by one and see what the 1.5 % is. So this is the problem. And Genesis has this, consumes a lot of time and effort to identify that particular information. So it's knowledge extraction use case. So how we tackled this is basically internally we can do certain things, but we need to collaborate. So we worked with external partners to see how much they're  experienced in this area and what is the cost benefit aspects to it, and what is the technical depth to it because that is also very much important for us to make sure that the relevancy is there. So we did all this and that is the first part to it. Do you have any follow up?

Nagaraja Srivatsan (26:45) No, I think it's a good start that you have millions of documents and you're trying to find almost like a needle in a haystack and you're getting an over-volume and search is actually just looking at keyword pad and giving outputs of everything and there's no context to it. There is nothing and therefore very expensive resource time is spent in trying to go through that. So it's a very fair business problem which is that across the board. So now. Tell me, interesting, how did you go about solving it?

Vinod Das (27:14) Yeah, so this is the catch, right? For two and a half years, I've been struggling to get this moving because there is a legal aspect, is a patent aspect, right? And a proprietary aspect. So you need to do the security assessment, both from internal and from an external provider, whoever you're going to nominate for, right? So this...security assessment goes with around 100 plus questions. It depends on the organization. The security will be based on what sort of data nature it is, what sort of information will be shared, what sort of key parameters are there within the document and where is it residing. So there's lots and lots of questions which will go with the legal. So that will be that. The vetting process took around six to nine months. So once you clear that, next you need to understand, what is the best fit team, as I mentioned. This team who was knowledgeable with security, with data science knowledge, with deep learning knowledge, et cetera, et cetera. So you need to get to that resource aspect. Then you should see internal articulation, whether my testers who are contributing to this data are available for giving the ground truth prompts for them for the implementers to succeed. So it's a multifaceted approach which is right now running. So yeah, we are moving in the right path, I hope so. And yeah, this is where we start.

Nagaraja Srivatsan (28:48) And what would be the ROI that you're going to go from these 200 documents to very specific outputs? Are you looking at efforts saved in each of those speedier queries? Because you're a very big person on business impact management. So what is that business impact you're trying to?

Vinod Das (29:06) I initially started with a problem statement, around 200 plus documents retrieval is there. So if we can reduce that to say five, after implementing this, then it's itself a great win for us. Because from 250 to five, five documents, we are reducing a huge effort. So we are incrementally looking for that, and we are progressing towards it.

Nagaraja Srivatsan (29:32) And as you're building this stuff out, this is your classic where you're assessing security, legal, proprietary, all of those things. But is that a magic model you're using, or is it just using, once you've done that, you're using your own proprietary, Chat GPT, or OpenAI models? Are you looking at small LLMs where you're building your own model? Walk me through, how did you make a build versus - I like, you have a framework, right? Build versus buy. So how are you evaluating a build versus buy in this situation?

Vinod Das (30:00) This is a small, small question to the tool selection framework also. So in this particular situation, we had the internal development build, which contributed to the success. And when you consider, OK, I'm building it, what are the changes which is required when you build a model? For example, we have our text embedding and other models contributing to that. For example, Cloud Sonnet, Cloud 3.7. 3.7 is not there in EU. It's Cloud 3.5 is the limitation which we have in European Union. So we are limited to one level behind. So understanding model is also important. So we had a combination of model to suit our use case needs, but we didn't have the biological model yet. That is one thing which is not available in the Cloud provider hosting. So we were nullified by the prompting aspect. We need to take our diversion towards additional prompting.

Nagaraja Srivatsan (31:14) That's a good point in it, and maybe we can just explore that in the last part of our session which is you know the whole thing around maker, shaper, taker you just made that very good stuff you know you didn't have a large Bio LLM, which is you know for you to build so you shaped and took it with prompting so walk me through how do you, you know, is your thinking - You're going to be doing more of making new LLMs. Are you going to take these things and shape them with more prompting, more Ras? Give me how do you approach that journey of maker, shaper, taker?

Vinod Das (31:49) When you start educating, the more you shape your data, it will be more effective. For example, Alpha Fold Pre. So we internally assisted on the beta side of things, and it didn't hold good for our data. So if you think in that angle, internal orchestration or internal small language model will be more effective. Because we are the knowledge base. Why not be bold? A proprietary model with our data. That is more powerful because we have the expertise which is required for building that. It can be a small, similar to Pi 3 or Pi 4. It's a small language model. So that is more powerful. When you think on a taker, we can take it. Say for example, if it is really effective, for example, Text Gemma, it has more than 22 tasks in which it is specialized on so we can observe it but when we really use our data and see the algorithmic aspect that's where the the challenges are okay because the predictability and the outcome, right should be reflecting on what we need in our therapeutics area. there are a lot of complexities in the maker, a taker, a shaper. It's a it's a fine-grained combination of how you use your strategies to utilize the marketplace as well as observe your efficiencies.

Nagaraja Srivatsan (33:16) Fantastic, Vinod. I think I took a lot from this session around A, your framework for evaluating the right use cases. I love the four quadrant framework. You also gave us a lot of thought around prompt and how you build prompt capability within an organization, both through training as well as prompt modeling. And then as you started your agentic AI journey, you've talked to a little bit about how that journey will influence it. So, Vinod, thank you so much for your time today. Really appreciate it. And really excited to have had you on the show. Thank you.

Vinod Das (33:52) Yeah, thanks, for having me here. And I sincerely appreciate it, and I wish you a great rest of the day.

Daniel Levine (34:00) Sri, this was a very nuts and bolts discussion. What did you think?

Nagaraja Srivatsan: I think Vinod gave us a very practical thought on how you build these different projects together. I mean, his whole framework on how he created the prompt infrastructure, the training of the team around prompts, the way in which the prompts became reusable - they're all key takeaways as people start to build gen.AI projects within that organization.  

Daniel Levine: The first use case you talked about was the case of protocol authoring. you asked about tangible outcomes, and he talked about improvements of 25 to 30%. I wonder what expectations life science executives may have who may not be that familiar with AI and look at it as some kind of a magic box of results someone would expect?  

Nagaraja Srivatsan: I think it depends on the context of use. Certain contexts of uses can actually give you 100, 200 percent, if not more, improvements. In this particular case, it's a very complicated change management program. Protocol authoring has usually been the holy grail of clinical development. Many people have started to enforce templates, and he said he used the Transcelerate template, which is a very useful template for what a protocol and protocol authoring should be. But with that said, there's creativity and flexibility depending on the protocol author. And he's also provided an infrastructure that you could be giving them the inputs, but also giving them the flexibility to change them. I think that's a very right way from a change management perspective. But I think there are certainly AI use cases where you can get even much more bigger productivity improvements.

Daniel Levine: You also asked about training. I think some people may think adding AI is just a matter of a plug and play addition, but how important is it for companies implementing AI to make sure they're adequately preparing their staffs?  

Nagaraja Srivatsan: I think it's very important, think prompting is a language in itself. You know, he talked about using an interview technique, STAR methodology to really shape your prompts, and the STAR methodology is describe the situation you want a large language model to respond to, tell it the tasks which it needs to follow very clearly, give it some of the activities it should do, and then finally describe how you want the results. I think it's a very structured way of learning and doing, and so it's very important for people to get trained in that model. It's both a mental model, it's also a model that the more you experiment, the more you get comfortable, and the more you can use and leverage these tools.  

Daniel Levine: You also asked about build versus buy with large language models. I imagine even if someone takes an off-the-shelf model, they would need to customize it and train it. Is that the case?  

Nagaraja Srivatsan: So we went back and forth in this discussion around maker, shaper, and taker. Makers are...where large companies are building these new large language models, whether it's Anthropic, OpenAI, Gemini from Google. These take lots of money, lots of GPUs, lots of effort. Shapers are where you're taking components of these large language models or open source models and making it contextual and building small language models for you. And that's where he was talking about the biology rated LLMs, which are much more tweaking towards the biology specifics. But as you start to take these functionality, he also talked about the practical way of using large language models, but using prompting and other techniques to make sure that it is doing what you want it to do, giving it a much more of a smaller context window, making sure that it's doing the tasks it's supposed to do rather than hallucinating across the board.  

Daniel Levine: I think the conversation really demystified some of the process that's necessary around all this. It was a great conversation and thanks as always.  

Nagaraja Srivatsan: Thank you so much. Really appreciate it.

Daniel Levine: Thanks again to our sponsor, Agilisium Labs. Life Sciences DNA is a bi-monthly podcast produced by the Levine Media Group with production support from Fullview Media. Be sure to follow us on your preferred podcast platform. Music for this podcast is provided courtesy of the Jonah Levine Collective. We'd love to hear from you. Pop us a note at danny at levinemediagroup.com. Life Sciences DNA, I'm Daniel Levine.

Thanks for joining us.

‍