Dr David Bergvinson: Well, welcome to another exciting edition of Grounded Intelligence. Now, last time we talked about benchmarking, and today we’re gonna be talking about localization. You might ask, “Well, what does that mean?” Well, you’ve come to the right spot. We’ve got the experts here from Athena Infonomics to help us go through the discussion paper that they put together on this very topic.
And so just as a brief introduction to provide some framing here, uh, localization is, is really, uh, using, uh, data and adapting local models to better serve communities, especially, uh, advisory services for farmers. And that localization is so important because it encaptures the context, making these recommendations more accessible, relevant, and ultimately actionable and empowering for farmers.
So, uh, with me [00:01:00] today is the Athena Infonomics team. We have Atamabar Chavan, we have Prashant Balraj, and Dr. Harsh Vats. And so welcome to the podcast. Uh, if you don’t mind, just if you could provide a very brief introduction of, you know, your role within Athena Infonomics, uh, that would be really helpful.
So, uh, Atamabar, why don’t you start?
Audumbar Chavan: Thank you, David. Thanks for having me here. It’s a pleasure to be a part of this podcast. My name is Gaurembat Chawan. Uh, I’m working at Atkins Economics in the AI and DPI services practice, and I’m leading the agriculture AI work on behalf of Atkins and also working on agriculture DPI advisory for our sub-national government clients.
Thank you so much.
Dr David Bergvinson: Thank you. Uh, Prashant, can you, uh, provide a brief intro, uh, to our audience of, of your role within Athena Infonomics?
Prasaanth Balraj: Hi, everyone. Uh, it’s a pleasure to be part of this [00:02:00] podcast. My name is Prashant Balraj. Uh, so I work as an associate director of AI and digital transformation at Athena Infonomics. Uh, over the past years, I’ve been w- designing and developing AI solutions in public sector, especially in healthcare, education, and agriculture.
So we have developed and deployed solutions into government integration platforms and portals primarily in India. And, uh, in my last stint at Wadhwani Institute of Artificial Intelligence Global, I was also part of their global team, uh, taking up solutions in Africa especially. And I’ve joined Athena very recently, just two weeks back.
Great to be part of
Dr David Bergvinson: Welcome. And Dr. Gotts
Dr Harsh Vats: Uh, hi. Uh, thank you, David, uh, for the, uh, introduction. Uh, so basically, uh, my name is Dr. Harshvardhan. I am presently working as a program manager, health in Athena Infonomics. So I have a half a decade of experience in public health. I worked with several government AI, [00:03:00] uh, programs with, uh, ministries. I was, uh, sev– working as a s- uh, sole liaison officers with, uh, several ministries to create, develop and deploy, uh, AI solutions, uh, across, uh, India.
So that’s all about it
Dr David Bergvinson: Wonderful. And so collectively, you bring a wealth of experience around implementation of artificial intelligence across a range of sectors. Uh, our podcast is focused on agriculture, especially advisory services, but we have much to learn, I think, from the other domains like, uh, finance, healthcare, uh, even governance instruments and the role that artificial intelligence can play.
So, uh, this topic of localization, hopefully we can draw on some of the lessons learned in other sectors as well as we go through our, our conversation. You know, one of the, uh, topics that was covered in the discussion paper was around governance and, uh, you know, localization as it pertains to, to governance.
[00:04:00] I’d like to, you know, given your experience in actual implementation of AI advisory services in particular, but just AI services, uh, especially through the public sector, y- you know, what are the real, uh, governance issues we need to sort of think about as we start developing, designing, and delivering, uh, AI-based services, uh, to society?
So, um, Atamabar, would you mind just starting us off with that response?
Audumbar Chavan: Yeah, sure. I think Prashant and Harsh are better placed, but with my limited experience, uh, particularly in context of governance, I would say, uh, there are a lot of issues that are plaguing, uh, the AI implementation across the public sector. The first and foremost is, you know, uh, who are– who own the data?
Who’s, uh, in, in the charge of the
duties to implement this system? So there are a lot of questions about the accountability [00:05:00] whenever system goes wrong. But, uh, at this moment, I would take a pause here, and I’ll come back to governance question in the context of agriculture.
Dr David Bergvinson: Okay. Well, may– uh, it’ll be great to get everyone’s perspective on it, given it’s so, it’s import-so important, the governance one. So Vasant, would you mind, uh, just giving your perspective on governance as we look at framing these solutions?
Prasaanth Balraj: Especially with respect to AI solutions, uh, governance is very important in terms of accountability. So once an AI solution is developed, like let’s assume that a particular company or an organization has come up with a AI solution and they have developed it with a particular grant. After benchmarking it and after taking it to the field, what is the frequency of updating the model?
Because we know that AI models has, depending upon the quality and depending upon time, uh, the d- models can degrade actually. So it’s very important to set up those cadences at what point of time one reviews the model’s [00:06:00] performance with respect to, to live data performance and who does that. And what are the new data sources in which these AI models needs to be again retrained or fine-tuned actually.
So the accountability of not only on the development or re- uh, refining of the model is needed, but also who takes the accountability and ownership of setting up these cadences, uh, taking it to validate, getting an external validation done, and owning the process is very important
Dr David Bergvinson: Yeah, well said. Uh, Dr. Vats, uh, what would, what’s your angle on the governance and, and its role in framing these solutions?
Dr Harsh Vats: Uh, thank you. First of all, brilliant question, and I agree with Prashant on this. So majorly, governance has an issue, uh, in relating to accountability itself. So basically, uh, if you go into the public health as well, if you, uh, create, uh, several AI models, after point of time, who will take an accountability at a ground level?
So you have to take [00:07:00] care of the contextualization in which spe-specifically in this term, we are calling it a localization. So, uh, basically, if I create a model at an India level, would it be applicable at a tier three city which does not have an access to the same, uh, environment, same, uh, internet and other, uh, possible combination which– in which environment you have created it?
So, uh, who is taking those responsibility? So, uh, once you deployed it, if you get out of it, there may be chances that the adaptability and accountability would fail after some time. So you have to make sure the system is sustainable when you are coming out of a system. So that needs to be taken into governance.
So actually at the end, it’s a will problem, not a capacity problem that has to be, uh, generated at a ground level.
Dr David Bergvinson: Excellent point. And, and that sort of takes me to the role of different levels of government. Um, I know, uh, the [00:08:00] Athena Economics team has worked very closely with several state governments in India, and I think you probably see differences between different states in, uh, assuming that responsibility and engagement.
Um, so I’d be interested to get your, uh, perspectives then on the role of state governments. Even though there’s a national mandate, let’s say, around, you know, AI for agriculture, uh, which is the case for, for, for India, but each state, uh, is sort of transposing that into its own context. Maybe we could talk a little bit about the role of these state governments or subnational governments that are, um, really t-taking on that ownership, responsibility, and sustainability of these services.
So, uh, maybe Dr. Vats, I’ll start with you, and we’ll back up the, the list here. Uh, you know, I know Andhra Pradesh, for example, is one state where you’ve worked very closely. Uh, can you relate some of the experiences there and how the state governments are taking on this role, uh, of localization?[00:09:00]
Dr Harsh Vats: Uh, yes. Uh, so basically localization depends, so, uh, specifically how state government would work as a missing layer. So generally, what happens is you create something at, for example, if you are creating any AI model, generally those are the open source model. You contextualize that model onto that specific state data.
So you have to make sure that the state person and specific those districts are involved who in which you want to deploy those solution. So once you create that, uh, AI solution, you want to evaluate. How do you evaluate is more important than even creating. So if you evaluate it properly, you would know the things that, that are actually getting answered is right, contextually verified or not.
Because if AI hallucinated, then who would be held responsible? There is no particular thought process on that. So, uh, in Andhra Pradesh as well, the thing that is helping us is that state government is also initiating those steps. [00:10:00] They are creating the accountability layer. So, uh, we create AI solutions.
They create an accountability layer, so the sustainability of a solution would increase exponentially. So that is a thought process because even with health and agriculture, both of which become a state matter, not a national. Uh, so national there is a mandate, but state is held responsible for it. So if the, for example, national government has given some mandate of, uh, 100 or 1,000 crore, then if state supplement that with their own, uh, network, then that network and that deployability increases exponentially
Dr David Bergvinson: Ex-excellent point, and, and, and spoken from practical experience, which is so valuable. Uh, Prashant, do you have any-anything to add to that, uh, around the role of state governments?
Prasaanth Balraj: Adding to Dr. Hush’s, uh, point with respect to to national and state level implementation, I think they play a very complementary role. For the national [00:11:00] governments and institutions, they create the policies, they create the DPI infrastructure layer, they create the platforms. So they create the interoperability of these, uh, solutions to be plugged in into different states, actually.
Whereas the state plays a very important role when it comes to data readiness, implementation, and, uh, training and capacity building of the users who are going to use the respective solutions, let it be frontline workers or let it be farmers. Uh, the state has a very important role in terms of getting it to the last mile people.
Making the content available in respective local language, whereas in the case of Andhra Pradesh, it is Telugu, whereas other states have their respective state languages. So make sure that if the LLMs are being implemented, making sure the dialect is in, intact with respective state’s implementation, whether the respective implementation layer post which, like, the structure or the organizational, uh, layer of the people who are going to use it could [00:12:00] vary from state to state, actually.
So that’s where the local implementation comes into picture. So it’s always a complementary layer and the communication between national leaders and state level leaders in terms of implementation of any digital platforms, let it be non-AI solutions also, it’s very important for when it comes to uptake of these solutions.
Dr David Bergvinson: Well, you mentioned one other layer, which is that last mile delivery layer. And so I’d like to talk about the human inter-intermediaries, uh, for the delivery of these services. So, Audambar, could you tell us a little bit more about the role of, you know, human extension, uh, workers play in the delivery of these services and the localization and building of trust?
Uh, so a, a few thoughts on, on those critically important human inter-intermediaries that, you know, this technology isn’t displacing, uh, extension workers, it’s complementing their relationships and ability to provide services.[00:13:00]
Audumbar Chavan: Absolutely, David. So in the context of agriculture, the role of human interval, which is very important because for, uh, for a very long time, the, uh, the farmers, particularly the smallholder farmers, they have been relying upon these extension workers for receiving an advice. And all of a sudden, in the last couple of years, we are talking about digital agriculture or AI-enabled agriculture.
But who are gonna the farmers? Who, who are at the frontline who deal with the farmers on day-to-day basis? Today, if you go to a smallholder farmer, so the, that trust has already been created by the extension workers.
And, uh, in the context of India, an average age of Indian farmer is, uh, 55. Some of them are not, uh, digital-savvy or tech-savvy. Uh, in the… So if we have, uh, an extension agents who are empowered, uh, through this, uh, digital literacy, and if they can take that advice to the farmer, uh, uh, this is very important.
And one [00:14:00] more thing in the context of India, what we have observed, uh, in across the Indian state that one of the trustworthy person in the Indian farm is a fellow farmer. Farmer is not gonna listen to any application. He will listen to a farmer who is a quite experimental progressive farmer, and that farmer will create a ripple effect that, that will be a force multiplier.
So I think, uh, no matter what, whatever, uh, no matter what, what could be the sophistication of model in the coming years, um, there is no replacement of the, be it an extension agent, the lead farmer, or there are many, uh, dozens of, uh, players who give the inputs to the farmers on day-to-day
basis, uh, which is, uh, absolutely an imperative for the AI systems to take an account for the human loop in the system.
Dr David Bergvinson: Excellent. Anyone else wants to add to that, the role of human inter-intermediaries?
Dr Harsh Vats: Um, uh, just a point, uh, which, uh, [00:15:00] complementing, uh, the Odamba’s point that, uh, basically we, uh, can’t think of AI solution as a standalone at a g- uh, last, uh, mile worker. It has to be a human because the dependency of a human-to-human connection is always there. And the chances of if AI doing some predictions or some, uh, processes wrong, there are a person who from the experience can override that decision.
So that is the same that we have followed with the WHO framework that, uh, human override shall always be there. Also, the important thing is we have to be also thinking about those people who does not have an access of an AI solutions. For example, women farmers, non-smartphone households, uh, remote, uh, rainfed, uh, areas where the access is not there.
In those areas, what would you do? Because, uh, you have to also think about those, uh, specifics. So if AI, rather than, um, thinking about it, AI shall complement it, not [00:16:00] to deplete them, but to complement their knowledge, uh, which would become easier with time. So, uh, for example, any, any, like if I am an anganwadi worker like, uh, at a remote area, I may not have an, uh, access, uh, to AI solutions.
But I could call someone up, uh, from my, uh, supervisors. She could guide me. Then I will pass the same information from my experience. There will be two persons’ experience merged with an AI. So it is a complementary thing. So that is something we have to think about the sustainability as well. So that is just a point.
Dr David Bergvinson: Yeah. You got an excellent point. I mean, it, it really leads us nicely into the topic of, you know, equity by design and making sure that, you know, we’re inclusive in, in the development and delivery of these services, either directly through digital or, as you point out, uh, through extension workers or lead farmers in the community.
Um, Dr. Vats, do you have anything to add on that equity by design component? [00:17:00] What are some of the key learnings, uh, from your work in different state governments in India on that point?
Dr Harsh Vats: Yes, absolutely. So the equity by design has to be very, uh, well thought out process. So, uh, uh, for-from some of the examples that I could give is that if you are planning any solutions regarding the malnutrition across India, you have to think it through that there are nearby different state. If you move from the north to south to east to west, the variation is very high.
So in the design perspective itself, you have to be planning that the data that is going into the AI model has to have those, uh, sample across. Along with that, you also have to see the prevalence of the adverse outcome has to be recorded, which is normally at– prevalent at a same population level. So all– if all those points are covered, uh, even, uh, when you are creating dashboard at the– for the final [00:18:00] visualization, you cannot, uh, judge a social impact just on some markers.
There has to be some understanding at a ground level. What are the value chains that you have actually made an impact on? Some are maybe qualitative, some are not be quantitative. So you have to think it through. Those are, uh, I think equity by design, as well as you have to also make sure, uh, what you are doing specifically from the design perspective, it has to be validated by some experts which are there in the field, and it has to be like a population proportion.
So, uh, whatever the issue is going to be on ground, it, uh, must be managed, uh, before even, uh, doing or training any models.
Dr David Bergvinson: Anumbar, you’ve already talked a little bit about, uh, you know, equity by design and approach with, with farmers. Um, so I’m gonna ask Pra- uh, Prasanth if he has anything to add on the equity by design?
Prasaanth Balraj: I think Dr. [00:19:00] Harshwardhan covered, like even during the time of model development, it is very important that you have representative of the data for where you’re going to deploy it. I want to talk a little bit more on the deployment aspects of the solution once it’s available, actually. So let’s assume that you create a solution that works online, actually.
But if your user base is going to be, uh, in tribal or rural areas where they don’t have access to internet, it is very important to create offline versions of those model, actually. Keep them into considerations. And also have other, uh, channels to make your solutions available in case if the smartphone penetration is not good enough, how are they going to access the services using simple phones, actually?
Like considering if it’s just a voice-based interference, actually. Uh, they can have it integrated into, uh, like WhatsApp channels or through simple IVR calls where they can call a toll-free number and can get a s- uh, like an [00:20:00] advisory thing from that particular chatbot, actually. So these are some channels that can be considered to cover the bases of the people who don’t have access to, uh, the respective AI solutions in low-resource settings.
It’s very important for the uptake of the solution.
Dr David Bergvinson: Those are all excellent points and, you know, especially in emerging markets where infrastructure may be weak or even absent, uh, for the delivery, digital delivery. Yeah, having these alternate channels is, is so important and, you know, different price points for technology. You know, you pointed out, you know, the, you know, sometimes we call them the dumb phones, but you know, the, the, the simpler phones,
Prasaanth Balraj: The feature phones,
Dr David Bergvinson: a smart-
Prasaanth Balraj: yeah
Dr David Bergvinson: Yeah, the feature phones.
Yeah, exactly. Uh, so yeah, all these are really important points as we think about, you know, designing inclusion, including the delivery channels.
Prasaanth Balraj: And from a AI developer standpoint of view, actually, I’m telling you, like, it’s very important to consider the cost of development of making your models available in those channels. Takes a lot of effort and, uh, [00:21:00] time to get those models out, actually. So when it comes to the final pricing of your services, we just cover only the actual cost that we are occurring.
We– A lot of, uh, people don’t account into the kind of effort which goes into make your model available in all these channels. That’s very, very important and a challenging thing in these emerging markets
Dr David Bergvinson: E- e- exactly. You know, so, you know, people don’t think about the full cost of these services for localization, and as you expand the reach of these services in reaching marginalized, uh, segments of society, those costs go up because you need specialized data, potentially even specialized models to serve those, uh, communities.
So, so how do we find the right balance between the cost of localization and reaching everyone, but still making it affordable and sustainable for governments or even private sector service providers? What’s the– How do we manage that tension? Um, uh, Atamabar, why don’t I [00:22:00] start with you on that one? Cost of localization
Audumbar Chavan: Absolutely, David. I think the cost of localization is one of the crucial, um, uh, important point to think about. Uh, earlier this year in the month of March, I did a field visit in Baramati, which is a sugarcane belt of Maharashtra, which is at the heart of sugarcane belt. And one NGO in collaboration with Microsoft is working on AI
agri pilot for farmers. But the cost of that particular pilot for the smallholder farmers is up to $100. It may sound cheaper to us, but for a smallholder farmer, that cost itself is, uh, very high. In that direction, I think, uh, the government system
has to invest. We have to invest AI as a public good instead of, you know, thinking only commercial players will take the lead and we will come out with only regulations. Like similarly, on the
lines of digital public infrastructure, if we [00:23:00] start treating AI as a public good, I think then, uh, we will make this model affordable for the smallholder farmers. Just touching upon the previous point, we were talking about the equity by design in the context of agriculture. We have two important seg-segments of the farmer.
One is a tenant farmer who does not necessarily own the land, but who is a cultivator of the land, who, who, who does the farming for a substance, subsistence or to sell the produce in the local market. And the other important segment of the farmers is women farmers. In academic social circle, we talk about the feminization of agriculture.
Women are at the forefront of agriculture. So when we think about the cost, I think the only way to solve the cost problem is there has to be some kind of substantial invest-investment by the state. But you know what happens whenever the grant gets over, uh, the project, the projects, they [00:24:00] stall at the pilot level.
So government going forward, they have to think it has to be a budgetary provision, I would say.
Dr David Bergvinson: Well, one of those costs is the, you know, data collection curation, uh, that’s driving these models. And as we get into more localized models, you know, those costs actually increase because it’s harder to actually collect, uh, data from, you know, less connected communities, let’s say. So, so how do we manage that component, uh, when we talk about costs of getting the right data, but, you know, still staying within budget?
And, and you bring up a really good point, Anumbar, about sustainability of this. After a project ends, who k- you– continues to, you know, curate and improve those data assets, uh, going forward? So, uh, Dr. Vats, uh, w- maybe you could give us a little bit of insight on data quality, the challenges that it presents, and maybe some of the solutions that, uh, have evolved in your engagement with state governments.
Dr Harsh Vats: Uh, yes. Uh, very nice question. [00:25:00] Uh, so, uh, if you think it very thoroughly, you have to be a very inquisitive and, uh, very thoughtful in finding those solutions. So the, uh, first of all, uh, which, uh, Odambo has already said that it has to be a very, uh, state has to be also responsible for that. But apart from that, from the ML angle, there is another thought that is contextualization.
So basically, what you would do is rather than training the complete model onto the specific state data, what I would take is I would, uh, get some proportional sampling, uh, so population proportional sampling from that area, and then contextualize that model. Just, uh, for example, if I ju- ha- uh, had just a validation layer using that data.
Let’s see, uh, how much result is skewed from, uh, skewed from the original results. Do we need to revalidate those? If not, then use the original AI model. If required, your, [00:26:00] uh, costing of the data collection would, uh, decrease substantially. S- because you have to very, uh, thoughtful. You have to remain in the budgets, and you have to, uh, also think about the local, uh, issues.
So, uh, what generally, uh, from the ML angle, what we does is, uh, use a contextualization, get some population proportion, then, uh, test that AI model. If there is an issue, then, uh, retrain the model. Uh, otherwise, uh, let it go. And after some time, the important thing which in the past, uh, question is as well, which, uh, Prashanth has pointed out, the important thing again is revalidation of a model after point of time.
So that is more important from, uh– So that has to be into the recurring line of, uh, recurring budget of the state or, uh, any specific, uh, solution which is deployed across that state has to be considered under that, like, uh, what we consider for the salaries. After six months, if you [00:27:00] revalidate everything, there is no issue that would, uh, remain there because, uh, you have to also consider the data drift and all those things.
Dr David Bergvinson: Yeah, excellent points. You know, I, I especially like the fact that it’s a evolving process, that you start with a model, uh, that you think serves the basic needs and, and that’s been developed through stratified sampling. But you’re coming back and verifying that, you know, these models are continuing to perform well, which leads me to ask the question, and we’ve touched on it already, is around the accountability and when those recommendations are incorrect or inaccessible, uh, or that, you know, that leads to a farmer making a, the wrong decision and, and actually losing money, um, by following these recommendations.
Can we explore, you know, what are the consequences of that and what are the systems in place or mechanisms to address those? And, uh, Dr. Vats, you point out that, you know, you could refine the model going forward, but what happens when, you know, this has already happened to a [00:28:00] farmer? How are those situations being handled currently in India?
Dr Harsh Vats: Yeah. So basically, uh, I’ve, uh, already told, uh, this specifically earlier as well. So there is a human-in-loop approach. So you generally always have a person towards the end of the loop of a prediction. For example, if I am predicting a malnutrition for any kid, there would be a, a Anganwadi worker who, who is actually on ground.
She would also judge from her experience as well, not working directly on the prediction itself, but, uh, using her experience. So we are complementing her experience with an AI solution. That goes across agriculture as well. For example, if I suggested one solution in, for example, if you have to put nitrogen or ni- uh, urea in, uh, soil, they would also, uh, uh, sug- like if I suggested that, uh, there would be a person on ground who would think it through as well, like before supplementing it, before giving a solution to other person.
[00:29:00] So you have to be very thoughtful of this. Uh, also,
Dr David Bergvinson: But are there, are there enough people to actually do that? Uh, do you have enough extension workers to do that? I mean, quite often these services are passed directly to farmers and, and what happens when they follow that, uh, recommendation directly from an advisory service that’s digitally, uh, delivered?
Dr Harsh Vats: Uh, so I totally agree with you. You cannot completely, uh, ignore the fact that, uh, the population is very large in those cases. Uh, what we do generally is, uh, we create the, uh, evaluation framework. So we evaluate the mo- models and provide, uh, the framework and the results. So for example, uh, I did a pilot across one district.
In that district, I evaluated that specific model, uh, in various, various places, just a population proportion. And if the issue arises, then I would re– uh, go and redesign or, uh, retrain the [00:30:00] model. If not, then it would be deployed. There may be chances we, uh, it, it is already like when we deploy anything, we will specifically s- uh, state it out to state as well that this model has this sensitivity specificity, and it– there may be a chance it, uh, the result can be, uh, uh, NPV and, uh, as, as, uh, NPV and PPV depend on, uh, the sensitivity and specificity.
Dr David Bergvinson: Yeah. I’m, I’m trying to tease out what happens when a farmer follows advice and they lose money. What is their recourse? Audimbar, have you, uh, had any experience with that in your work with state governments? How are they handling that?
Audumbar Chavan: So at this point in time, first of all, we have to keep in mind that agriculture is a very sensitive issue. There is a political economy angle to it. If advice goes wrong and there is a dip in the production, it is not necessarily affecting the individual farmer, it is affecting the nation’s food security.
And if there will be demand-supply mismatch, it will also impact on the [00:31:00] inflation. So there are these macroeconomic, uh, implications if most of the farmers will start implementing the advice, AI advice and advice
to ground. But from the individual farmer’s perspective, uh, at this stage in India, we are at a very nascent stage. Uh, we haven’t come across any example, and that’s where I think the role of agro-economists, that human in the loop, that’s the topic that we are currently again and again we are referring to.
Even in Baramati, there are these
agronomists who are checking whether that advice or the data or the insights generated by, uh, the AI system is right or not. And then we are passing that advice to the farmer. Um, otherwise, if farmers start adopting these advices, then there are going to be huge consequences, uh, both from the individual farmers as well as, you know, the entire macroeconomic picture for the nation.
Dr David Bergvinson: Great.
Dr Harsh Vats: Also, uh, just [00:32:00] to add a compliment on Odamba’s point, uh, so what it does generally is, for example, if it is an, uh, LLM. So there is also in evaluation framework, there is one model in which what we do is, uh, so once the results are out, it has to be expert vetted. So within those, uh, LLMs result, uh, just below that you will see expert verified.
If it is not expert verified and ex– uh, like there are s- uh, different, different questions, uh, then in– if it is not expert veri-, it says AI generated. So, uh, the person who’s reading also would have a context that this is AI generated, not completely expert verified. So that would also compliment and help us, uh, so that the, uh, things that in ground shall not go wrong.
Dr David Bergvinson: Excellent. Excellent. Well, um, I’d like to just sort of wrap up with some closing thoughts or reflections from each of you. Uh, I also invite our audience to click on the link below to access the [00:33:00] localization paper that has been developed by the AgX AI community. Uh, so you’ll find that resource, uh, here to dive into this important topic in a lot more detail than what we covered, uh, today.
But this is just to whet your appetite for the topic and to, to dive deeper. So Atamabar, what would be your, your key takeaway from the localization, uh, paper, the discussion paper that you’ve developed?
Audumbar Chavan: Yeah, in our paper, we have called that
out very clearly that localization is a governance challenge. Uh, it’s not java– It’s not just about having the right language or right agronomic
practice. Ultimately, the responsibility bears down on the institutions which are, you know, having the ownership of validating the content and checking the accuracy. So just to sum up, I would not say this is just– localization is not a add-on feature. Uh, it is a continuous commitment, uh, for ensuring that the farmers develop the trust on the system.
So with that, I would like to end here.[00:34:00]
Dr David Bergvinson: I’m glad you mentioned trust. I, I’m, I’m timing how long it takes us to get to that word, so thank you for
Audumbar Chavan: NARM©
Dr David Bergvinson: that up. Uh, Prashanth, uh, what’s sort of your key takeaways from the discussion paper?
Prasaanth Balraj: One thing we have ta– we talked about a lot about data technology model, etcetera. But when it comes to localization, it’s ultimately the people, I mean, the farmers over here who are going to use the AI or digital solutions here. So the thing is that ultimately if your solution is not going to fit into their regular workflows, if they’re not going to see value out of your solution, they’re not going to use it actually.
So when you asked about, like, what’s gonna happen if the AI model throws an unverified response or a response that affects their response? They’re not going to use it. And they’re a closed network, so while farmer A from one particular district is going to communicate to farmer B, so the word is going to travel and they’re not gonna use it actually.
So it’s very important as AI developers we see what is the right fit, understand from [00:35:00] farmer’s point of view. Do not think only from a technological standpoint of view or a platform point of view to solve that problem. Understand from their point of view and see if your solution is going to be a fit or not.
Because AI alone cannot solve all the problems. AI is one of the ways to solve and mimic human intelligence and make it easier for the farmers actually. So that’s very important. And second thing, when it comes to, uh, equitable– equity when we are talking about it, from a technical standpoint of view as AI developers also, we need to be responsible in terms of localization actually.
We don’t need to do primary data collection for all localization activity. For example, if I have to build a computer vision model to identify different pests, uh, for a tomato or a coffee, uh, crop, there are a lot of open source datasets that is available that’s from India or Ethiopia. Let’s assume that if it is available in Ethiopia from– I can use that, and most of the model architecture and the refining is going to be the [00:36:00] same.
So I just need to, uh, collect a little bit variety or a representative dataset from India to match that particular thing and reuse that code. So I can save a lot of time and resource when it comes to approaching the model development itself before I hit the field and find the right contextualization before hitting the field.
So that’s two things to remember. Keep in mind whom we are designing it for, and also as, uh, AI developers, we need to find ways to make the resources used and put up in a right way and save cost and
Dr David Bergvinson: Yeah, human-centered design. Human-centered design is so important for this especially. And Dr. Vatz, uh, your closing thoughts, reflections on the discussion paper for localization. What was your big takeaway?
Dr Harsh Vats: Yeah, big takeaway from, uh, both the angles which, uh, Prashant has already complimented. It’s important to understand the localization because at the end, those are the people that are actually going to utilize your solution. So you have to be very thoughtful about them. [00:37:00] And very important is that you have to evaluate very properly, because if both of this would not go as rightly as you stated out, then the chances of acceptability as well as adaptation would definitely go to hay
Dr David Bergvinson: Wonderful. Well, gentlemen, thank you so much, uh, for this interesting discussion on localization and the other elements that touch on localization that position these services for success. So really appreciate your time. Uh, again, I invite our viewers to, uh, click on the link below to access the discussion paper to explore this topic in more detail.
So Adam Bar-Prasanth and Dr. Vats, thank you so much for your contributions today, and we look forward to continuing to evolve this conversation because it is a rapidly evolving technology, and we need to learn and adapt as we implement. So really appreciate your expert perspectives on this today
Audumbar Chavan: Thank you so much, David
Dr Harsh Vats: Thank you, David[00:38:00]
Dr David Bergvinson: Okay. Well, that I think is a wrap. So I would just invite you to, uh, join the community of AGX AI
