Uh So hi, everyone. And uh I'm speaking from India. My name is Kalna Bansal. Um They're trying to talk a little bit about what we're trying to do with cognitive intelligence on talent platforms.I work in a SAS company which is a front unicorn uh that is trying to scale an enterprise product, uh which revolves around talent data. And this is about our journey with talent data, Maria. The next slide, please. Uh We've seen some interesting stuff that has been happening with data in the talent space or in the hr technology space. We have moved over the last uh 40 years from uh actually measuring or calibrating data to actually trying to anticipate the requirements for employees or the workforce.

And that has created a need for a whole new kind of cognitive process that can actually look at um how employees trace their life cycle journey. What's actually happening to them in the course of their careers and people are no longer in single careers, but in multiple careers and uh what's actually happening to the skills uh the experience and the unique parts that each of them is following and that's been our journey so far.

Uh Maria the next slide, please. So there's a whole lot that has been happening uh around the space of employee experience, talent. Uh uh linkedin is a prime example of something that has been trying to do uh something with the organized workforce. Uh There are certain elements that actually comprise or create a talent profile that includes your achievements, your certifications, your education, uh what do you do with projects? And that's a very, very interesting topic from an it perspective because technology companies look at projects as a very holistic uh environment or ecosystem in itself your job history. Because no two job histories seem to be entirely alike uh language, which creates a whole lot of modalities when it comes to actually uh moving talent from one location to the other. And then of course, the universal currency that has become the topic of debate, which we shall discuss in a little bit more detail, which is skills and this has impacted not just one but multiple journeys across the platform uh ranging from what we call our talent strategy to what is actually happening on performance management.

Because people are increasingly saying we do not want to go online and actually put in data on performance. How are there lot of ways of doing that using some kind of cognitive intelligence? What's happening with? Uh how do you plan successes or careers? How do you automate the career paths for individuals and therefore, what do you do in terms of compensation for specific skill sets or for specific roles or projects that are being undertaken, Maria. So some of the very, very interesting things that we have uh toggled with and a whole lot of it. Uh everyone knows Ted GP D and everyone has been talking about C GP D but Ted GP D is nothing but a transformer model which is actually an eval. Uh It's an evolution or an existing model called B. And we were trying to deal with this larger problem saying in a country like India, we have about 10 million applicants who are active on the organized job front out of these 10 million people. And most of them come from good schools and colleges. How do you identify or search and match the right kind of person for a particular role? How do you therefore lend them a job descriptions that make sense uh for their unique roles? How do you man uh manage a huge candidate data pool that is about 10 million records? And how do you create a wide talent search across your social networks to land the right kind of person?

And this has been a problem with uh one of our leading banks has actually uh thrown to us. Uh Essentially, we have deal with about 100,000 job postings a year which leads to about uh a million plus resumes which hit the job recruiter. And how do you therefore conduct search and match on such scale. And uh we had two or three rather challenges as we were doing this one is we noticed that when you look at different models in artificial intelligence, which try to do a match of a job description and uh resume, uh we noticed that things like the competencies, the education all need to be extracted and that there is no standardized dictionary across the world in terms of competencies and skills, different people label those skills differently.

So some company calls it problem management and other company calls it problem solving. Uh competencies tend to be called either negotiation skills or influencing. And it becomes extremely difficult to land a very, very accurate search. We also realize that uh talent profiles that are technical in nature uh write their resume is very different from the ones who are non technical in nature. And which simply meant for us that a model that is as big as this, which actually renders a search and match using. Uh we use a bit of cosine similarity and a little bit of elastic search. But um the complexity of the model is that we really needed to get very domain specific in terms of addressing at scale, a very accurate search and match that is not just based on keyword, but based on context and also to land a recommended job description or even uh passive candidates from the past pool that could be linked to a particular job in the easiest possible manner.

So that was our big problem. Uh Our last problem was about talent search because there were companies that came back to us and said, uh we want somebody who has done construction of airports of a particular size in a particular geography. Now, when you're trying to do such a large search, you face three challenges. One is uh what are the exact nuances or context of the search? Second, what is the kind of time that the system will require to do that search? And if you notice if you search for something like this on a link, if you search for something like this, on a job portal, you'll probably land the match in less than a minute. Uh being able to land a match very accurately in a very low response time has been the biggest challenge uh challenge as far as we can see uh on this whole experiment that we have conducted Maria. Um the other places that we have started using A I and uh I, I get scared to say I know anymore because a lot of this is actually now auto ML uh But on the career path side, uh we have started looking at uh drawing up what we call unique career trajectories for employees basis where the skills that they have uh and their skills profile basis, what they aspire to do to also recommend some career paths and also to allow for a very, very hyper personalized trajectory.

So there may be some guys who just want to go horizontal. There may be guys who want to go vertical and we have provided for both. And of course then to enable the search and match, you will actually have a very very complex talent marketplace in place which does auto recommendation uh curates your different job and po project postings and also promotes DN I by pushing the female candidates for example to the top in the case that you actually have two candidates of the same caliber or instance.

Uh Likewise, what happens with the individual development plans because this is an interesting one when you have large enterprises, an individual in an enterprise is actually receiving feedback. Uh In fact, if you ask me too many sources of feedback and that actually becomes a blockage to the whole process because uh the individual very often does not know what to focus on and how to look at his development plan. So one of the uh nice uh usages of A I has been to auto curate the ID P basis three different input points. One is what is the current skills profile of the employee? Second, what are the employee aspirations? And third, what is it that he needs to deliver on his key metrics in the current year? This is that the auto ID P actually writes the individual development and of the employee and puts the top three priorities so that the employee is not really spreading his energy across multiple different sources. And likewise on the succession planning where you could configure talent pools using skills which are available and also do something like a comparator basket across multiple candidates to see who is more suited for a particular position at a point of time.

Yeah, Maria now, uh where is this being used? And we, we are an enterprise technology and we compete with recipe oracle workday. Uh The usage of something like this happens across the life cycle on the employee platform from workforce planning, which is how do you create the right mix of talent, whether it's part time, full time gig contract, uh remote working. Uh how do you therefore decide which positions to push externally, which to push internally either on your talent marketplace or through succession management? How do you therefore calibrate your performance management and reward systems aligned with the kind of skills that are required?

And therefore how do you create a career management framework that works? Yes, Maria. Uh so if I put it in a very contextualized manner, I think uh where we are on the journey and what do we see happening? Right. Uh while we are saving a lot of time and resources and I am, I'm sure all of you who have used linkedin would have noticed at various points of time that while it is always quick, it may not really be the most efficient recommendation engine that you have had. Uh I have myself encountered a lot of challenges with linkedin because it keeps uh sending me job postings for an intern when I have 20 plus years of experience. So there's there is evidently a mismatch between the kind of matching algorithm that is happening at the back and what's put out there on the profile. And uh so saving time and resources on one hand, but being able to optimize what's happening with multiple resumes or multiple profiles that go into the ecosystem using smart engines and smart models or even layering multiple models of A I one over the other very, very important.

And at the same time, uh the domain adaptation of such a model needs to be extremely accurate because like I said before, if you use a general language model or a LLM for doing something like this, as you may seem to speak an entirely different language, which has nothing to do with uh what is there in the English language.

And training the model to understand the specific nuances of a resume in a particular industry or sector has actually led to a very, very uh fine tuned, very accurate results and also extremely automated candidate screening and a very very efficient filtering mechanism on candidates.

It's also been able to bring the model to the level where we could actually uh tell the recruiter or the internal manager where the gaps are in the client's profile or the candidate's profile, whether it is skills, whether it is company, whether it is expertise or whether it is proficiency, uh the critical parts of it and what's actually creating the huge challenge is of course, being able to extract keywords uh very, very accurately.

Um If you look at the scale of an open air and if you look at the scale of the organized workforce, um you, you come up with some interesting challenges because uh when we came down to keyword extraction and to some part, we did use open air to do the keyword extraction. And we found that it was not very accurate. We actually found manual tagging by freelancers who actually understand the domain far more accurate but much lower. Uh So while in open air attacks, about 600 resumes a day, uh something like a freelancer doing this manually on a per day basis would have somewhere around 100 but the 100 tended to be more accurate than the 600 that an open A I or A chat. Uh Any of those language models can actually extract for you. Uh Also in terms of understanding which keywords are crucial to extract and which is not crucial to uh extract. And that was a challenge that we encountered through the journey. Uh Mario and at the backbone of all this, which is I think the most important part of this whole story. And journey for us is what do you do with something called a skill ontology. And uh linkedin talks about skills, everybody talks about skills. Uh companies are talking about skills. Uh The challenge with skills is that uh you know, there is no real, real repository of skills where in the world we have. So we have on it, all of them are skilled libraries.

But uh the challenge when you're trying to find specific kinds of talent or specific kinds of nuances of talent is that you don't need the skill repository, you need the relationship among the skills because uh something as simple as driving. If I say driving in India, I'm talking about somebody who drives with the right hand. If you're saying driving in USA, you're looking at somebody who's left hand. And now if you're just putting a post, saying I want a driver, I could land up millions of records. But if I say I want a left-handed driver who's actually driven on mountain is terrain. The probability of landing somebody that actually currently within a context uh becomes much uh lower if there is a very, very nuanced way of calibrating the different kinds of skills and the context of the skills. So what we have tried to do at the back end is to represent the skills and nodes and look at the relationship between different skills. And when we say relationships in an ontology, there are different ways of classifying it and um that has helped us to create a very, very nuanced search.

Now, the only complexity that you have here is that as and when you keep uh skills, keep evolving relationships between skills, keep changing. And also the complexity of the implementation keeps changing. Um It's a very, very interesting piece of work because uh we really did not find any uh model or any library that has an AY that is so well trained and which renders it in graphical mode. So we actually had to create the ontology using uh LLM model and then render it using a Neo four J to actually render the process. Um Maria and like I said, therefore, in terms of concluding uh remarks, uh personalization and optimization for individual journeys at scale is something that uh EIML actually lands beautifully, large amounts of data can be analyzed and multiple data points can be uh collaborated uh in terms of causation to actually identify which ones actually impact your career path.

Uh The journeys on employee or talent management are therefore becoming smarter and leading to higher job satisfaction. But uh along with that comes the caveat that uh the more the ambiguity that we're having around job roles, the kind of skills that people uh and in terms of expectations in terms of delivery, uh the nuancing on these models is becoming more and more complex. And I'm going to rest my discussion here and leave it open for the last two minutes for any questions or insights that might come up here or any inputs that anyone wants to share. Um If there are no questions, then I'm done. Thank you very much for hearing me out and I wish you all a very, very happy day. Thank you so much. Thank you.