Leveraging GenAI to Accelerate Innovation

[Music] uh thanks for joining us um I know this is the last session of the day and I don't want uh me to be the person standing between you and your evening plans um I'm actually stepping in for chanoy uh of mankind farmer he could not join us because of some Visa issues so uh what I'm going to talk about is what we did with man Pharma the challenges that we faced why we actually took a gen route in the first place and what are our learnings uh things to do and things not to do

so these are what I want to talk about to give you a brief about Mankind Pharma uh it is uh India is one of the largest in terms of volume both in ter terms of prescription as well as number of drugs that they actually sell in the country they are into multiple uh areas around uh some kind of diseases like chronic disease acute disease as well as some of the consumer products as well so they're all over the place uh from a uh organization perspective they have like 15,000 field reps on the ground uh talking

uh to millions of doctors across the country Gathering a lot of information uh presenting their case and finding out what needs to be done so it's a very complex operation now let me give you a sense in terms of philosophy Mankind Pharma wants to uh support the tier three tier four as well as the villages in the country which means they have three important challenges one is affordability second is quality and third is accessibility and why is this so important one is they are actually competing with generics in terms of price however from uh a

quality as well as mind share perspective they are actually competing with global Giants for similar products so they have they are in a unique dilemma a tough Balancing Act to do so that's their philosophy of what it is now when we started to work with Mankind Pharma we saw three important things one is in terms of scale the scale was huge both in terms of transaction as well as in terms of the volumes of uh the various aspects that go into it second is complexity complexity in terms of languages dialects the way it uh works

the diseases in that specific uh place so it's kind of complex and the third important aspect is around the cost since they are competing with generics it was very essential for them that they ensure that any solution that they built is really cost effective for delivering their products uh what I wanted to do is I want to hear from Chiran jooy and what are the challenges that they're facing and the various initiatives that they have started to uh work on hi we at mankind farma started our journey with data braks about a year back after

thorough evaluation of the platform we decided to move all our Enterprise data into Data bricks Lakehouse architecture as we go through the process of building this foundational layer we have already identified quite a few use cases that is going to impact our 15,000 plus medical Representatives who serve close to a million doctors in India with 1,000 plus skus or products within our portfolio all of this will drive data democratization so that each and every employee within Mankind Pharma can take better and more productive decisions to give you a couple of AI and Genia use cases

that we are already developing as pcc's one is we are creating a natural language querying engine for our Frontline sales people so that they can ask a simple English language question to get answers and insights from internal as well as enter external data similarly we are using jna to create pack new package Design This We Believe are only the first steps towards the AI and gen transformation that Mankind Pharma is going to take in the next many years to come thank you thanks for that so before I get into the details of the solution he

talked about three important aspects um I want to give you a sense in terms of what the entire organization is going through they have thousand skus and it's very very difficult for the field reps to understand uh all the Thousand skus the information around that so typically when you actually look at them when they go to the doctor they need to pitch to the doctor in terms of uh What that particular product is some of the packaging details the various statistics around that that's very very critical when they have a discussion with the doctor so

remembering all of them or trying to find that is been a big pain for all of them the second important aspect is once the prescription happens they have to quickly figure out the suppliers side or delivery side of the information in terms of hey where is the inventory which uh chemist has what amount of inventory where is the shipping information these are very very critical because if the doctor prescribes something and the uh medicine is not available in that area you typically find them uh giving an alternative and thereby losing a business so it was

too it was very important for them to look at what I call the supplier side or the delivery side of uh information and finally from a language perspective if you are actually working with 15,000 uh people all of them are not fully English trained everybody has a slang you you would see me speaking in a very different English compared to most of you but in India if you look at the complexity we have 26 languages written and spoken okay so they are very very unique to each other add thousand plus dialects to it so it's

now you can understand the complexity the second important information is when they are conversing with their doctors or their suppliers or their chemist there they need to converse in their local language which makes this entire uh complexity uh like multi fold let me give you an example uh let me take a very ordinary example not the medical one for example if I need what I say sugar cane juice it's a simple two-letter word in India people might come and say hey I need ganka juice ganaka is a Hindi word juice is an English word you

see the combination of it unless an llm understands both the language and knows when to split it is very very difficult so these are the kind of challenges the Reps were going through so when you look at some of these U operation challenges regulatory challenges it might look very simple and St straightforward but when I add the uh various aspects that I talked about language visibility as well as the uh complexity it makes the life of the field rep as well as the organization very very difficult and when we started to talk about gen to

the customer the executive leadership there was very clear they did not want to spend a lot of money on it because of their philosophy second they were very very protective of their IP they did not want anything to go out so that was the second important thing and the third important aspect then when we were talking to them was hey can I actually make it simple for to the rep on the field that is the Big Goal that they all carried and though it looked very easy uh it is very very difficult one the second

thing that we worked on with them was was around how do I do packaging of the product with thousands of skus they had a large team me team of people who design the product who label the product whether it's internal or external set of people spend a lot of time and money and they wanted to see can we find a better way in terms of doing this which makes the makes it more cost effective and quicker as well and finally if everybody wanted voice and I gave you an example of a couple of words that

we use how do you use voice from the various uh uh stakeholders but get the desired result that they want and I'll explain to you some of these aspects in detail so let me take the first one so firstly we said hey Genie room seems to be a good place for you to start one in terms of understanding your own information so it was fairly straightforward and when we started off this discussion gen room was in private preview and it was only available in the US and moving data out of India was a no no

because of all the uh regulations that we had so we actually uh took a sample data we changed a lot of context we moved it uh to a US data center to have them test everything now when they tested this they had three three observation one is in terms of accuracy the accuracy was 70% which I thought was good from a natural language processing but the customer said this is not acceptable okay the second thing was when we moved data here a lot of context was missing because we took specific data we said hey let's

use the data for our POC and testing the context was missing so it was very difficult for their team to say hey this works well or this is the right answer or wrong answer and the third important aspect was the uh Missing of the various uh semantic information or information beyond what we've supplied to the uh geni room which created a lot of problems now when we started this discussion and we had this multiple rounds the customer was all almost ready to say hey this is not the right solution for us let's move on now

here I want to actually thank our engineering team they actually took each and every feedback uh they made those changes that was required because there was some challenges with our Genie room and how text to SQL was getting created second we uh they also enabled this in India region which made our life super uh helpful and third is now we had the entire data I mean they had their own data to test it out when they did this their results were very very encouraging the organization was happy uh the leadership seems to be super happy

with it so in the first round we said instead of opening this up for the entire organization let's open this for only the CEO office CE of office where they can actually ask straightforward questions hey sales by SKU where are the diseases spread uh which areas are we seeing better traction uh these are simple questions that they were uh able to get in now as stage one is complete we are working on stage two which is helping reps to actually ask similar questions where they can actually pick up a particular package or ask for a

particular SKU and say hey give me uh everything about this sko in terms of what this sko is what are the uh what are the critical information what are the statistical or data points that I need to use to talk to the doctor and convince them that this is the product that they need to prescribe so that's the second stage that uh we are looking at um it's been an interesting journey I would say when we had smaller data set which was more structured we had better results but now we started to open to larger

data sets um our learning here is don't start opening this to all your data sets it is um unless you are 100% sure on what that data set is and you have control on it um the results are not great now with that said as we got this learning we'll now be slowly opening up data set after dat data set on the geni room uh we have already enabled like six uh teams right now and we are further enabling to uh a larger team as we talk about so uh as a gen product we said

hey what's the simplest thing to do use Genie room to move forward that's step one the second thing I talked about the field reps finding out what about a product and other related issues now this is exactly where we had to use our gen I rag architecture what did we do so we said hey why don't you actually use the information around your various products um inest that and start asking questions so the first one to jump in here was the CEO office you know what they did they actually ingested the investors report as well

as uh the calls transcript of all of their organization as well as their competitors organization and they use that to understand the various aspects or various questions that they have in it was very important because this gave them an understanding in terms of what medicine was required where is the information and how to pitch it to a doctor so it's super important that they started to look at and the next step that they have started to do right now here is to create the first drafts of regulatory reports using llm now uh when you do

regulatory reporting as in any country it is not one for the organization you would have to do one for SKU level uh a set of uh different products so the amount of regulatory reports that a organization had to file was so large that using um the rag architecture that we created they've been able to use this to generate the first drafts of the report which can further be modified and presented this actually reduced a lot of time to the CFO as well as the uh regulatory office that they had so that's the second use case

that we are working on third uh important aspect that we are we were working on was um around the um the package itself when we look at the package there lot of small components around this and you typically have questions the doctor ask a lot of questions in around that and it's very difficult for uh the field to remember all of it so we gave them an option where they can actually use voice to ask those questions and then uh get the relevant answers so it was a combination of rag plus voice where we converted

the voice to text and using text we started to ask questions and got the relevant results so it's work in progress as I speak but we are seeing some good traction there now here is an example that uh we used in Mankind Pharma around different use cases the organization was very clear hey one the answers have to come out quickly the second it has to be factual meaning they wanted to drill down and look at the almost like I used to always call it's like the lineage of data you know I want to know where

is this information coming so it was very important for us to present that hey this is the answer that uh you are asking for however here are the pages from which we have picked up so when you click on the link that you see here it actually um takes you to the exact page from where we are picking up the information and summarizing them this gives this gave the entire organization um more uh they were uh more happy one second is they also KN knew that since it was factual information they could present it uh

to a lodger audience beyond their own so that is an interesting take that we did uh from a rag architecture perspective now the Second Use case that we did was to redesign the logo and the need here was to create a 3D design design um and how do you incorporate your logo the various uh regulatory information as well as the product contents that you typically uh give so you can imagine doing this for th000 eses it becomes so uh hard so what we did is we worked with them uh on multiple things so let me

give you a sense of picture yes yeah so there was significant prompt engering work we did we actually work with Deli to start off with and you see how the uh packaging came frankly this was not acceptable to the regulatory authorities it was too good and Mo and mostly it looked like a 2d print rather than a 3D print so in a package when you really look at it and I did not realize this still working with them that 2D is not acceptable it had to be a 3D uh picture so um we we also

worked with them on stable diffusion the learning that we have is at this point of time the 2D images are really good it shows a lot of things however uh there is some more significant work that we need to do to create 3D uh pictures one second is to tone down the kind of um pictures that you see on the uh packages because it looks too flashy for a medical product that is very important so that is some work that we are currently doing uh the learning here is uh clean can we actually have more

data points around medicine medical industry to ensure that when you generate these kind of images text to image you get better outcomes rather than outcomes which looks great on picture but cannot be used practically so what they did is they actually Ed this picture on all the digital ads that they wanted to provide and they had a very different picture on the product so it's uh work in progress we are saying hey if somebody is not able to correlate your digital picture to your actual picture on the product it becomes slightly difficult for you to

sell so that's the second use case that we have worked on and uh finally the Field Force have to submit Daily Report now one of the uh when we met their CEO he said hey rajes I like the information that I'm getting with from them in terms of how many doctors they have seen how much of medicine how many prescriptions have happened what is the feedback and all those but I'm not getting a sense in terms of is it the real picture or what is the mood on the ground so what the CEO wanted to

understand was he wanted to understand how is the rep feeling when he's actually giving this uh info is he feeling great or is he actually just giving the data for uh from a checklist perspective saying I need to submit the report daily can I do it so the project that we are working on is why don't you actually submit your daily report using voice and we want to do couple of things one is to analyze the voice to see what is the voice saying is he happy is he confident is he feeling hey this is

not going to work this is uh a couple of aspects they want to look second the entire when we talk about voice the data we get is in very unstructured way how do we structure the information that we get in such a way that can be used for reporting purposes that's the second uh part of it and the third part of it is once we get this how do you collate all of them now all of the reps are not going to talk in English they going to talk in their respective languages so now it

adds to all the complexity that we talked about and this is an use case that we are currently working on with multiple models uh we are seeing some good traction here but there's still a lot of work that we need uh we need to do so from a Mankind Pharma submission of daily report PS seem to be like a oneliner but if you actually peel the onion you will see a lot of things that the organization wants to know and do so that they are able to keep the Field Force happy and uh also get

the desired information that they're looking for so it's a very unique uh gen use case that I would say it's not like a point use case but it's like multiple uh llms that we need to stitch together to make it work so that's the third important use case that we are working with them probably if he come next year I would give a better update on what we learned what was working what is not working but at this time it's too early okay so just to summarize I wanted to talk about hey natural language is

super exciting every and this is the quote from the customer itself they are currently evalua ating uh they started with multiple organizations can we do it at the Enterprise scale that's what they learn second as I said the image generation is at a very early stage we need to revisit that can we find different ways to use it um to make it more uh product friendly slash regulatory friendly and finally the design and processing of audio analytics is super complex um we've been working on various aspects of it um as we go forward uh in

the next few months you will see us uh maybe presenting a white paper on how we uh went about doing this and how to achieve what you want so those were the aspects that I wanted to uh talk about great thank you very much [Music]