ACM ByteCast - Kate Kallot - Episode 70

Episode Date: May 29, 2025

In this episode of ACM ByteCast, our special guest host Scott Hanselman (of The Hanselminutes Podcast) welcomes Kate Kallot, founder and CEO of Amini, an impact-driven AI company based in Nairobi, which focuses on the critical issue of data scarcity in Africa and its implications for AI development. Before Amini, her career spanned leadership positions at global tech companies, including NVIDIA, where she led global developer relations and expansion into emerging markets, and Arm, where she was a pivotal figure in the Tiny Machine Learning (TinyML) movement. At Intel, she led the development of the world’s first AI development kit in a USB form factor, the Neural Compute Stick, bringing computer vision and Al to IoT and edge devices to millions for the first time. Kate is a recognized expert and influencer in the AI field, advising international organizations and governments on the potential and challenges of AI for good. Her work has been recognized by TIME’s 100 Most Influential People in AI, the World Economic Forum as a Tech Pioneer, and One Young World as Entrepreneur of the Year 2024. A trusted voice in global AI policy and digital equity, Kate serves as Vice Chair of the ICC Global Environmental and Energy Commission and is a member of EY’s Global AI Advisory Council. In the interview, Kate explains the barriers to AI adoption in Africa, stemming from challenges with digital and environmental data infrastructure. She shares her work collecting and validating data in key areas such as climate and agriculture through state-of-the-art technologies and partnerships with private companies, using a bottom-up approach. Kate and Scott also talk about Amini’s commitment to open source and community collaboration in areas such as geospatial data science, and the global applications of Amini’s work in Africa to other geographies with similar characteristics. We want to hear from you!

This is ACM Bytecast, a podcast series from the Association for Computing Machinery, the world's largest education and scientific computing society. We talk to researchers, practitioners, and innovators who are at the intersection of computing research and practice. They share their experiences, the lessons they've learned, and their own visions for the future of computing. I'm your host today, Scott Hanselman. Hi, I'm Scott Hanselman. Hi, I'm Scott Hanselman.

This is another episode of Hansel Minutes in association with the ACM Bytecast. Today, I'm chatting with founder and CEO of Amini, Kate Kallon. How are you? I'm good. How are you? I'm very well. Thank you very much for spending time with us today. You're joining us from your offices in Nairobi, is that right? Indeed, here in Nairobi, Kenya. Yeah. And amini.ai is the website and one of the things that you come right out and say on the home page, you say data is the biggest barrier to AI adoption in the global south. That is the thesis

here that we're working on. Why? Why are we missing data? Why is the Global South data scarce? One of the things you have to remember is that when you look at the digital divide, you still have 2.6 billion people that are unconnected in the world. And basically, unconnected means invisible. And most of those people are based in the Global South, which means they don't have internet connectivity, they're not online, so they're not represented by the data sets

that are being generated today to basically train AI models. So when I moved from New York City to Kenya, I just had an idea with my team, which was to build an AI company for Africa. But then as data scientists, we start building and you realize, well, there is no data, so what do we do now? So we thought that if we were running into that issue, we were actually the rest of the ecosystem as well. So we decided to pivot towards building

the data infrastructure for Africa and the global South. So we know that the corpus that AIs are trained on is wide and vast, but it's also biased towards English. It's biased towards the Northern Hemisphere. What kind of data are we missing? Because I can think of what I think. I can guess what I think some of the data is, depending on whether you're talking about data for a GPT, data for a chatbot, or whether it be more specific data. It's actually affecting many different types of data. So you would think about

data in terms of languages, for example. You have more than 2,500 languages in Africa. A lot of those are only spoken. They're not even written, right? So how are you going to go and find the datasets to be able to capture those languages, those dialects, that are only spoken by very, very few communities across the continent? The second one would be data in terms of environmental data, which is the first one, the first big, big, big challenge that we decided to tackle. When you think about, for example, satellite constellation, geospatial data, we have an

abundance of data in the global north. In the US, in Europe, you're able to access weather forecasts very easily on your phone. But when it comes to Africa, actually, the continent has only one-eighth of the minimum density required of a meteorological station. So how are you even able to predict weather and understand when there'll be an extreme rainfall, when there will be floods, when there will be drought? It makes it very, very difficult to access high quality weather data on the continent. So when you think about that data scarcity is not just one modality,

it's actually multi-modality and multi-industries. Think even about culture. A lot of our culture is not online neither. It's very, very difficult to digitize that culture and understand how the continent has evolved over the past couple of decades. So we've decided to focus on environmental, but it is a challenge that's affecting many different types of data, and there is a big gap to close here.

So when you say focus on environmental, is the issue satellite coverage? Do we solve that with drones flying over hectares of land? Like, how do you get this data? Because you're coming from behind in this race. You know, it would be very expensive to fly drones all over Africa. So that would be quite impossible. But one of the things to mention

is that we're using satellite imagery. So that's the first data modality that we started to work with. What we realized is that a lot of the constellations, the satellite constellations that have been deployed in the past are actually, when you start looking down to Africa and you start going into areas that are less densely populated, Africa becomes dark. There is a lot of work that needs to happen where you couple geospatial imagery with machine learning and AI to be able to really truly understand and extract insights out of this imagery.

We're using satellite or geospatial data as the first modality, but then we're also working with companies, startups, private sector companies who have a lot of data because we're talking about data scarcity, but it is also information and accessibility. Sometimes the data is there, but people don't know what to do with it. So we realized that we needed to put a lot of effort and made a conscious choice in developing our own ecosystem of data partners

who are fitting this data. Their data alone wouldn't mean anything, but their data coupled with other types of data, geospatial data, that's when it becomes very, very interesting to fusion all these different data types and see what are the source of insights that we can extract. And you call out what you call ground truth sources. And can you explain that you get a bunch of data,

but you don't know if it's true or not? What is a ground truth source, and why is that alignment so important? So when you look at satellite imagery, it's extremely biased, right? It's the reality of the sky. If you look down and you look at the ground, usually there's some sort of calibration and validation

that needs to happen between what you can see from the sky and what you can actually measure from the ground. So you actually need to have, just like in computer vision, you need to have ground truth when you work with satellite imagery. So what we do is that a lot of our data partners would go and, for example, collect soil samples, measure yields when you think about crop and farming, and we're using this to calibrate a lot of what we see from satellites.

The satellites that you're pulling this information from, are they just seeing the visible spectrum? Are they seeing information that might be stored but not being used? Are you squeezing, like squeezing the fruit of these satellites for information that we didn't even know we had? Yeah, we are. So even if you look at the open-source satellite from NASA and the European Space Agencies, a lot of them are actually giving you a resolution of 10 meters. What we're doing, for example, is that we're able to use augmented techniques to do super resolution and look at and go down to a meter resolution so we can see even more of what's happening on the ground.

One of the things to understand is also the different context. When you think about doing satellite analysis on the US, for example, especially if you think about farming, it's large farms, monocrops. It's quite easy to use our open source imagery and you don't have to do a lot of fine tuning with the image. When you look at Africa, it's actually one by one square kilometer farms, multiple crops, it's not as structured and it's not as big. It's very, very difficult to go down the pixel level and really understand what is in a pixel. My team has spent a lot of time building models that are actually taking this raw image and adding a lot of machine learning models that allow us to understand things like

lens use, lens cover, do things like crop masking, being able to detect pest and disease. Like a lot of things that you would see are micro, micro, micro details on the environment. We've been now able to combine both AI and data science with geospatial imagery to be able to really understand and fine tune what we can see from the sky. Now, I want to ask a dumb question and forgive me. This might be a myth or a stereotype, but you're the perfect person to answer it. I have heard as a person who lives in the Northern Hemisphere that like, well, you know,

there's cloud cover, there's jungles, there's things you can't see, like it's impossible to see into the dark continent. I think that that's a narrative that we are told. Can you squeeze more information out of these satellites by multiple passes? Is there in fact cloud cover? Is there a flagged places that you can't see into? Or do you have techniques that make that nonsense?

We have techniques that make that nonsense. So it's true there's cloud cover. If you look at Nigeria and Brazil, those two countries actually are very similar in terms of vegetation. And they have a lot of cloud cover. So what we realized is that we had to do techniques such as cloud masking.

We actually had to build our own version of Google Earth Engine because we needed to layer multiple different satellites to be able to get the insights that we needed. So our team spent so much time working on our geo package which is basically our geo processing engine that allows us to pick from all these different satellites. So we're working with different satellites, we're working with multispectral as well. So you would look at like satellite that I have a spectral as well, and being able to look

at a satellite that our SAR as synthetic aperture radar satellite. So being able to ingest those those different resolution different modalities of satellites, we've been able to really squeeze in the maximum we could from open source imagery and run our analytics as we can. So this has been almost a year and a half in the making, our geopackaged life is fully running. That's what we're using for all our analytics in Geospatial.

But we also know that Geospatial is not the only modality that you can use. So we've also spent some time working on ingesting drone imagery because you have a lot of drone companies, not all across the continent, but in some specific pockets, being also able to ingest tabular data or even data from IoT sensors. A lot of people have deployed IoT sensors. You have startups who are doing electric bike charging. All their electric bike chargers actually have a small weather station in there. So we're able to capture a lot of that data because they have sensors that can measure

rainfall temperature. Okay, you have a small weather station, can we use your data? So this is the type of little hacks that we've been able to do. And we've been extremely community driven and building that community around us to help us really strengthen our data sets, which then strengthens our outputs on the machine learning side. That idea that there are all of these electric bikes just sitting around there and they just happen to have a weather station, it makes total sense. But that scratches a part of my brain I didn't know that I had. That's such a clever, clever hack.

Yeah, it is, it is. You use this term cloud masking, and this is like a well-known thing in the satellite data ingress kind of universe. The idea that you've got cloud shadows, you've got clouds themselves, and then you want to eliminate them to ultimately get a cloud-free image, right? Is that the goal? A complete, perfect cloud-free image of the entire continent, down to the meter? So, yeah, so cloud-free image, not of the entire continent, but you would look especially, so we don't, when we start our analysis, we actually start at the micro level.

So we will start with a specific country. We took a bit of a reverse approach because we realized that a lot of companies who are working with geospatial data actually start from the global approach. They start from the macro approach. That's what they use from a satellite imagery standpoint, but also that's how they train their models. We tested a lot of models that are global weather models, global satellite imagery model, global foundation models for earth science. And what we realized is that when you're trying to look at downstream tasks in Africa, the

accuracy drops massively. So my team decided to take a very contrarian approach and we said, okay, let's start small. So we'll start with one region in Kenya and then expand to the whole country of Kenya and then start expanding to other countries, the region and then the continent. So we took a very opposite approach and we started understanding that there was a lot of micro details down to the pixel

that we actually needed to understand if we wanted to have an analysis that was anywhere near like a good level of accuracy on the continent. So when we even started training our own foundation model, we realized that we needed to start with one country first and then expand to more country and fine tune with a lot of ground truth data. And even in insert,

cause a lot of countries in Africa don't even have the same climate, the same vegetation, like everything is very, very different. From east to west, it's the complete polar opposite. East doesn't have much cloud masking, west has quite a lot. So cloud cover, sorry, and west does have quite a lot. So we decided that we needed to really regionalize the way we approaching our analysisize the way we're approaching our analysis and the way we're approaching building foundation model to just understand what's happening on the ground and then calibrate this with some of the global models that exist out there. That has been a journey that our data science team has undertook over the past two years.

We didn't start and said we're going to train a foundation model for a science for Africa. We actually felt that we could use some of the ones that are available, like we didn't start and said we're going to train a foundation model for a science for Africa. We actually felt that we could use some of the ones that are available, but we realized that the accuracy was not good. So that's why we decided to take that opposite approach and really do things like the continent is doing, like developers of the continent are doing, which is more bottom up than tops down. And with this, we've been able to reach some very decent and

acceptable on our terms level of accuracy, and we're now using this in our workflows. Now, you've mentioned open source a number of times, and I've seen on GitHub, like the Sentinel Hub cloud detector in Python and all kinds of things that I, as a person from the outside, can look at, can participate in. How much of the things that you're doing are proprietary foundation models that are private, your own IP, and how much of it are open source models that people can get involved in and help?

So I'm a big proponent of open source. So I've spent a lot of my career working on open source software, open source platforms as well. And with our team, we decided that we would release our foundation model to the community. So it's actually been released last month on Hugging Face, the first version. Now we're training V2.

V0 is on Hugging Face, V1 is being trained right now, and we'll continue training and fine tuning and releasing this to the community because we want people to contribute. We're not doing this for us. We're actually solving a challenge that we think we need more developers out there to support us to do. And that's why we actually we made the conscious decision to release that

with a research blog or our research and our we were accepted. This has been our first poster accepted to ICLR. So this in a two years company, we're extremely proud to now starting to contribute to the research, but also contribute to the open source community and we will continue doing so. That's amazing. So then, this is again a business type of question. You're not a non-profit, right? You're not an NGO, but you are committed to open source.

How do you find that balance? So there is a balance between the research that we're doing that we think can have an impact on our communities, including the local developer communities, and there is a balance on serving our customers. So if you think about what our customers are buying from us is mostly insights. So your global companies who are exporting commodities, whether it's agricultural or mineral out of Africa,

and you need more data on the first mile. So you would come to Amini for that. Or your government and you're trying to really change the way you're looking at infrastructure, you would also come to us because you want to understand, like for the first time, what are the actual resources that you have in your country? So we really, this is basically where we are generating revenue is between global private

sector companies and local government. The one thing we decided we will not do as a company is to make any of our communities paid. So developers working with us are not paying, we're doing a lot of competition on Zindi. So it's one of the platforms, a bit like Kaggle, but very focused on Africa and emerging economies. So we're working a lot with Zindi to help train developers on geospatial data

and understanding how that specific modality works, because it's not something that a lot of data scientists have touched in the past, data scientists and machine learning engineers, and it's much more complicated than computer vision. We are trying to support our ecosystem to understand how to work with this specific type of data, but also supporting them with providing them some data sets so they can build models that are adapted to their countries and their communities. Africa is large and we're targeting not just Africa but also the global south. So now we're working in the Caribbean, we're starting working in Southeast Asia as well. So there is space for everyone and we have a big emphasis and we want to contribute to the community. So that's something

we'll continue to do. So we've been finding that balance between the research, the community, but at the same time, making sure we're able to generate revenue and pay the bills. Very cool. Now, you say that your custom AI models are tailored not just for the global south, but also for industries. We've been talking about agriculture. You've got this model that is primarily focused on Africa.

But as you start to go into Southeast Asia, as you start to do something like Brazil, are those models applicable because of the southern hemisphere or is the agricultural ecosystem so different that you're gonna need a Southeast Asia model or a South Asia model that's different? So they're actually applicable because of the very similar climatic conditions

and topographies and population of farming or mining communities. So when we started, we only were focused on Africa, but all of a sudden we get people asking us, can you do Brazil? Can you do Colombia? Can you do the Caribbean? Can you do Nepal? And we're like, well, we can try. And once we started digging into, so one of the process that we have when we approach a country or an area is that we start splitting it in agroecological zones. Rather than splitting it in administrative level, so county, ward, or district, you have to split it in different zones that have the exact same agroecological conditions. So when we started doing this for other countries, we realized that a lot of it, if not 70%,

was extremely similar to what we had already done in Africa. So what we needed was more ground truth to be able to fine tune those models so we can expand the coverage and the use case. So that's why we spend a lot of time in building that foundation layer. So if you look at our foundation model, it can do things like land use, land cover. We're also looking into similarity search using embedding. So a lot of it is applicable to different regions with similar topographies and similar climates. I'm thinking again about the bicycles with their weather stations.

And I'm remembering a couple of years ago, my iPhone popped up and said that it had a barometric pressure system and Apple was asking if it wanted access to it, presumably to fix their weather prediction, which makes me realize that there's a ton of Android phones running around in the global south. That's all, they're all data,

they're all weather sensors, they're all, but they're locked up, it's locked up behind the device itself on the edge, or if Google has it, they're not letting us know. Is that true? Are there little tiny weather stations walking all over the Global South? Yeah, it depends on the sensors they have

and it depends as well if they have connectivity, but that could be the case, yes. I'm just thinking there's a lot of information out there that is not being utilized to your point. Now, are you the only company that is doing this? Is it just that the Northern Hemisphere and Europe and America have not turned Sauron's eye towards the South? Could someone swoop in and cause them any trouble if they decided to think that this is a problem that they need to solve? I mean, right now, we are the only company doing this, not on the agriculture side or the weather side. For us, again, we're data infrastructure companies.

So environmental is the first step of data we decided to tackle. Then we'll start expanding. We've already started expanding to more than this. So for us, it's really how can we go and close that data and compute gap on the continent and then give access to others to actually build solutions. I want to build an ecosystem. I don't want to build a product. I don't want to build 10, 100s of models

for different applications. I want others to do so. I wanna make sure that they have the right data sets, that they have access potentially to the right level of compute to do so. So we'll give them the tool, we'll build a kitchen that builds the tools for everybody else, and then they'll go and build.

So when you think about that positioning, definitely we are the only ones. And if you wanna a comparison of a very similar company in a very different industry, though, that has done the same thing, you would look at, for example, the Palantir back more than 10 years ago, when they were sitting in government's rooms, black rooms and digitizing data for the government and building things for for for the defense and the military industry. We're not in

the defense and military industry, definitely not, but Africa is where the US government was at that time. We're still not capable of capitalizing on our data. We still don't really understand how to unlock the full value of our data. We're all talking about clip frogging, AI. Okay, yeah, but there are steps to go through first. We are trying to walk our ecosystem down this path to becoming AI ready so then we can go and all start adopting cloud across all our workflows and across the entire continent,

but also reach a level of digitization where everybody can say, okay, Africa is truly lip frog and now you have LLMs available in multiple African languages, even the ones that are the least represented and such. And there is an abundance of data that we've been able to capture, and we've been able to really even the playing field for Africa, where now it's easy to access weather data, it's easy to even use Google Maps. Let me tell you,

it's still a challenge here. You have places where you get lost all the time. So a lot of these very simple things that we take for granted in the global north, we hope that by solving that data gap, we actually will be able to enable those things to happen on the continent. ACM Bytecast is available on Apple Podcasts, Google Podcasts, Podbean, Spotify, Stitcher, and TuneIn. If you're enjoying this episode, please do subscribe and leave us a review on your favorite platform. I like that you said that you're building an ecosystem and not a product, and it's worth noting for folks that may not be familiar with your work that you ran the developer relations arm of Nvidia and focused on emerging areas.

So you're a community builder and have been one for a long time. You worked at the TinyML Foundation. How do you are you doing? Like it's not really Amini's job, but I'm imagining, how do you go and build this data infrastructure company? But I'm assuming if you still have that developer relations bone, that you want to do hackathons and teach people GIS. Everyone's focused on GPTs and generating text. That's the hot thing right now. But you want to teach the young people how to work with this data. Are you doing anything like that?

Are you working with the local communities? We are. That program I mentioned with Zindi, we're actually running hackathons and we're doing webinaries. So I'm not the only dev rel in my team. I have a very big bias for dev rels. So a lot of my teams are actually community builders and that's what we love doing. So we spend a lot of time going like doing workshops. We work with a lot of research centers, universities where we go and present

about the company and do workshops on geospatial and GIS. We do with Zindi, we've been doing a couple of different hackathons. The latest one that was launched is a crop detection challenge for Côte d'Ivoire. So differentiating, for example, cacao from coconut trees, so being able to really teach those developers how to use geospatial data. So we've been still doing a lot of DevRel, and it's been part of our DNA as a company. I hope that as we grow,

we'll be able to do more of that. That's fantastic. I know that in the past, in the past 10 years, everyone has been saying that we need more developers, we need more programmers, and lots of companies from the Northern Hemisphere have been running all over Africa trying to say we need to teach more people how to be developers, but we need more ML engineers, right?

We need more machine learning experts on the continent that have the ability to work with this kind of data. Yeah, we do. And if you think about one of the biggest wealth of the continent is actually our use. It's the youngest continent in the world, and it will continue to increase. So how do you actually make sure you provide them the right tools for them to be able to really build a future that's different than what our parents and our grandparents have built up until now. I truly believe that it's going to be via technology and I've already been doing so.

They are so smart. A lot of them would learn machine learning and data science through non-traditional password institutions. They'll go to hackathon, to bootcamp, and watch YouTube, do Coursera classes. They would not go to school to learn computer science. They'll be self-taught data scientists and machine learning engineers. And for me, that shows how hungry this use is, hungry for new things, for challenges, for access to tools and opportunities.

And I hope we'll be able to be a catalyst for that. That's lovely. This is an infrastructure company, but there are infrastructure challenges that can't be ignored on the continent, depending on what countries that you're in. Now you've picked Nairobi as your home base. Have you bumped into energy sustainability or infrastructure issues as you've tried to do this work?

Because I'm presuming you wanna do the work on the continent. You don't wanna ship the data to Delaware and back. You prefer to do that on compute. That's in Africa. Yeah, we do have a couple of data science workstations, this whole supercomputer farm here in our office. So we've had to be very, very creative about how we power that. We have a UPS, an alternative power source, because you still have power challenges.

So one of my hopes for this podcast is that the power doesn't cut all of a sudden, because it still happens. Especially, it's been raining quite a lot the past few weeks, so there have been a few power cuts. Once the power is gone, then you have a generator, but it takes a couple of minutes to come up. There are still deep challenges on the continent. On top of power and electricity, you would add connectivity. There are places, a lot of places across the continent, especially in rural places where you still don't have access to internet, right? So these are challenges we've had to work around, but we've been very hacky in a way to make sure that we're able to power and continue to develop a lot of our workflows.

So now we have our supercomputers that are still running. Sometimes I come back to the office and I see, don't touch foundation model training. So yeah, it's one of the things we've had to work with, but also it's one of the things that gives us that little, like very different mindset than companies in the global north, in the northern hemisphere, you know, because we know how to build with our local context in mind. And this is one of the things that you can take for us. So if you think about a team that is well placed to really solve for that in Africa is not going to be a team that comes from outside. It's going to be a team that was born from within. Yeah. Very scrappy, very focused, very hungry. Yeah, I love it.

Exactly. So what is the... You know, they always talk about inventors having a eureka moment where they say, we did it, we did it. I'm curious in the last couple of years as you've been building these models, was there a moment that sticks in your head where you did something or the team did something and they called you? They said, come in, come in, we've cracked it.

And then you got to see something or answer a question that was unanswered before. So we have one of the first employees of NASA Arvest in the company. He is like extreme, we call him the professor, he's extreme geospatial engineer. And he was telling me about, when I first started, I was not a geospatial expert, background in data science and machine learning. So for me, it was new for a lot of us, it was new, because it was a new type of modality we had to work with. But he was telling me about all these challenges he's experienced with the likes of GE over the other geospatial platforms over the past couple of years.

And he's like, listen, this thing is really hard to use. The onboarding is really hard. There's a lot of, it lacks flexibility. It would be great if we can build our own. And I told him, I think we can. And he said, I don't know. You know, it's like, we're not, how do you go about building your own geopackaged geospatial processing engine?

Said, I'm pretty sure we can figure it out. It was last year when I got a call and he called and he said, Kate, we did it. For someone who had been in the field for so long, knew so much, and was just dreaming about something he could give to fellow students, researchers, that was much more easy to use than all the tools out there, and can do the same, if not even more, than what those tools do. Coming to the realization that he was actually capable to build his own, that for me was a really proud moment.

It happened at this time. It happened when the data science team launched their first foundation model, the first version of the foundation model, like those things that I call them the kids, my engineers, these kids didn't know they could actually achieve, you know, they're like this, we're usually consumers of technology, we usually use anything that's out there, we don't build our own. But just seeing how this helps shift their mindset around, actually, I don't need someone to do it for me. I can do it by myself and I can do it even better. I can do it adapted for exactly what I need,

was a really proud moment. That was my haha moment. And I think I'm going to have more as we continue to come up with new tools. It's been such a joy chatting with you. Your enthusiasm for the possibilities of the continent is absolutely infectious, and I really appreciate you taking the time to chat with us today. Thank you for having me. That was great. We have been chatting with founder and CEO of Amini,

a purpose-built data infrastructure for the Global South, talking to Kate Calot, founder and CEO of Amini, a purpose-built data infrastructure for the Global South. Talking to Kate Calod, founder and CEO, absolutely fantastic work that they're doing. You can check them out at amini.ai, A-M-I-N-I.ai. This has been another episode of Hansel Minutes, in association with the ACM Bytecast. We'll see you again next week. ACM Bitecast is a production of the Association for Computing Machinery's Practitioner Board.

To learn more about ACM and its activities, visit acm.org. For more information about this and other episodes, please do visit our website at learning.acm.org slash Bytecast. That's B-Y-T-E-C-A-S-T. Learning.acm.org slash Bytecast.