Utu Interview

There has been a considerable amount of movement in the startup ecosystem in the African continent the past few years. More solutions are coming up with more creativity and disruptive edge. Just like the first world countries, African youths are also tapping into Artificial Intelligence to make their products more efficient in terms of time, cost and as a result, increase productivity by a great deal. 

As part of our A4AI webinar series, we recently had an interview with Bastian Blankenburg, Co-founder and CTO of UTU Technologies, one of the most prominent AI startups in Kenya.

Read the full interview script…

Who are you and how did you get started with UTU Technologies?

I’m Bastian Blankenburg. I’m a co-founder of Kenyan immersion startup called Utu technologies. I’m a PhD in computer science. German originally grew up in Germany, went to school, then to university in a town called sub looking, which is not very well known. It’s in the southwest of Germany. I started working at the German Research Center for artificial intelligence as a researcher for a couple of years, and then started doing a PhD there as well. That was the topic of Game Theory and using that for multi-agent systems, which are a distributed system paradigm where we assume that we will have many independent parties working together over a network and each one trying to achieve some profits. I concern myself a lot with game theory, uncertainty mechanisms like fuzzy reasoning or probabilistic reasoning, financial risk measures, also, privacy preservation, so how can we work together in a networked world and achieve optimal results. With unknown parties without disclosing everything that we might want to keep a private topic that also gets more relevant.

After doing that, I went to the industry and started working for a company in Berlin called IVU, which is making public transport planning software. That was an exciting experience, just from the perspective of learning software for industry works. I was a project engineer for some time working in places like Saudi Arabia, Italy, UK, and Ireland. Then again, I switched to the development departments and became a programmer and then I moved to Kenya because of private developments. Since my wife is Kenyan and she was here, we then decided to move here because I had, of course, visited a couple of times and seeing that there’s this awesome tech community in Nairobi, who are genuinely trying to do great stuff. Meaning that, from my perspective, coming from Berlin, where everybody is solving first world problems, such as, order your pizza with one click less or think things like that. So I find it very inspiring, coming here and seeing your people who are trying to solve real problems and making life better for farmers and things like that. Then I met my co-founders, with whom we first started a company called Mara Moja, which is a taxi app and we timidly evolved into co-founding Utu Technologies.

Would you say you are one of the worst world leaders in the work that you’re doing? Explain why

Well, we certainly want to be, but we are not quite there yet. So when solving we started that taxi company, we figured out that in Kenya and similar places, we don’t have only the problem that Uber and the taxis are solving, which is, how do I find any taxi because being around CBD here at night or whenever, and not just jump into any cab as you want to be sure that it’s okay. So basically the problem is finding a trusted taxi and not just any taxi and therefore, we developed this mechanism where we offer people different drivers and then we show them whether the drivers are liked by any one of their friends or family or generally known people. So we started asking people to sign up with Facebook and it was access to their phonebook to figure out who knows whom. And then we learned when we mentioned this to investors and others that, yeah, that’s required. A lot of different sectors not only for a taxi, but anything where you would ask friends or trusted people for recommendation. In case you need a mechanic, you will go and ask your friend with the same make of car as yours. So that’s why we founded UTU as the technology company and focus on this trust recommendation mechanism by itself. So we are now offering this as what we call trust infrastructure for the on-demand and sharing economies. But any online marketplace can be served by that mechanism. And in the course of that we also have them developing algorithms, for example, to identify the most relevant recommenders because when we started, we were a bit sceptical, like okay, how many people do we get signed up that know each other or know a lot of people on the platform. So you get quite a fast-growing social graph pretty quickly. Just even when you start your first couple of hundred users, you can probably assume that some of them will actually know each other and so that’s pretty cool. Over time, you learn of course also about connections between them, and then you add more sources where you’re getting the social links to show them, and then you learn more about that as well and so on. Gradually then you get to the problem that not do I know anybody who can recommend something, but you have maybe like 20, possible recommenders and then you don’t want to show that to the user on a small screen because that would clutter the user experience. Then you have to identify what are the most relevant recommenders here by looking at the context. Like for example, for the auto mechanic, the context is the maker of your car, and maybe your preference of price was quality and then find somebody that you know, who you trust, and particularly trust for this particular decision. So its similarity matching some contexts and others, and that’s where the whole AI comes in. When you ask, how are you, leaders? Well, I mean, we are leaders in the sense that I think, as far as I’m aware, we are the only ones building a genuine purpose trust platform like that. Some other companies who build trust as a product or They’ve made some similar mechanisms, but it’s restricted to their own vertical as we started with a taxi up rather than offering this generally. And so as I think, in that sense we are. There’s another sense where maybe we might be I’m not sure. There are a lot of companies, particularly in the West, or in Asia also that are using big data analysis and using machine learning and data science for solving problems where they have a lot of data and that’s not always the case for us. So when we start onboarding a new platform, we might not start with a lot of data that we have and we can’t just have just deep learning. And then we first need a massive marketplace with millions of users before we can show anything. So that’s not our approach. Our approach is more like we are using some classical methods, AI methods. Like similarity matching, like cosine similarities, clustering, things like that, to, to work with smaller data sets at first, and then only adds machine learning over time when the system grows, and we get more data.

Why are you using AI and the challenges of using AI specifically in the African continent?

In principle, there is, of course, a lot of data available about almost everything. The question is just how accessible is that data. So, for example, I mentioned that in the beginning, when our platform started, then they might also be startups, and they might not have a lot of data. So that’s, that’s one thing. But the other thing is, you might have a platform that has data and maybe a lot of data because they have been in the field for some time for years. And we also have such clients who don’t have that data online. So that data exists like in written books, like transaction books and now, so far, they didn’t have too much pain using it this way, because they were not scaling anything new. But now they try to do this and then all the data needs to be brought online or digitised in the first case. And that can be difficult. I mean, you can, of course, hire a lot of people who would just start typing out all the data. You can also try to use OCR technologies. The machine learning policy has made significant progress over the past couple of years as well. So that’s worth doing Or trying. It’s not always that great anyway, because some of this handwritten data as well and messages, so inconsistency is how the format is between different entities, you have contradictions within the one book itself. Then you have a lot of things where people make notes that might be obvious to somebody who’s in the domain which has domain knowledge but for general that might not be obvious at all. And even for people like us who are not necessarily domain experts in our client’s domains, but then also be challenging. So the data intake digitisation, and that was the first hurdle. Then another one is you hear about talent shortage a lot, but I’m sure there’s a lot of Africans generally who are really into machine learning. We have a large young community in Nairobi but also in other places like Addis Ababa and others that are on the continent where people are getting into this, and we have some awesome junior developers also from Akiva chicks, for example. When it comes to senior experienced people, that’s of course where it gets a bit thinner and of which a lot are from if they are from Africa, who go to study in some western universities. Unfortunately, our head of data science Dr Alex Mwai could not join us here. I had asked him to, to talk more about the data specific stuff because that’s more his domain rather than mine. And then head of machine learning Brain Muhia, who is also very well known in the tech community in Nairobi. So these are fantastic people who we just happen to come across, and we’re able to hire because they liked to be part of what we are trying to build.


What are your expectations from A4AI, the audience, Government, Investors, so that you can move forward with the solutions you have and scale up and become a big leading company in the world?

That’s a big one. The Governments can do a lot in terms of regulation and making it not too hard specifically for AI startups to come up and Roll. It’s often the case that deep tech startups, rather than just applying tech startups, need to figure out the business case and find the product-market fit and possibly even test out whether the product that would fit the market is feasible to develop at all. For example, the Kenyan government’s introducing this digital transaction tax is just not a good thing. Because you’re not paying for profit, you’re paying for any transaction and that can kill a lot of efforts instantly, because the beginning will simply not be profitable. And then you still have to pay for your losses to the government. So that’s not a great idea. In terms of other regulations, the task forces that have been done around the continent regarding the blockchain and AI system in Kenya is a good thing if something tangible comes out of it. We are still basically seeing what they are now doing with the results. Well, first of all, they created a sandbox for FinTech startups and some things there which work partially for some, not for all. I’m not an expert in regulations but I just feel that some rules change and that makes it harder or good for us. The governments can also do a lot of making it more accessible And look at that, from this perspective, building up the industry, rather than how they can extract something from an industry that is very innocent, and very small but can become massive. So the government shouldn’t try to extract stuff from there. The other thing people should generally do, I can only encourage young people and computer science people specifically to take up AI. But I would also advertise to not only look at the currently still hype, machine learning stuff but also at the more classical fields of AI like logic. It is because even though machine learning has made the most visible advances say I mean, other examples of the video were driven by machine learning advances. So definitely, there’s a lot of glamour and hype there right now. But at the end of the day, it also has a lot of problems. For example, explainable decisions by the highest and solution in the realm of machine learning have been some research advances, but nothing really practical. Yes. And, and also, transferring machine learning models from one use case to another is also a highly researched topic and advances are being made, but it’s also totally not solved. We see this very much because UTU has this approach that we are working with pilot clients to develop models for the first. Then trying to extrapolate the general methods from there, so we do a lot of feature engineering, even abstract feature engineering where we say here we feature engineer this and for that client, we feature engineer that so what’s the everyday things that we did there and can we automate that. So we are also working towards some sort of meta-learning things. And for this stuff, it is essential to know your basics about abstractions and algorithms, classical algorithms, logic and things like that. So definitely for people wanting to get into the field, starting with the machine learning stuff that’s great to these courses, but also look at bits beyond that machine learning into other AI fields. 

Then about investors, and it is a big topic. That’s not only about AI, but generally investing in startups here. And then there’s also a discussion going on. Why, in places like Kenya, for example, startups, with white founding teams, as UTU also get more investment from foreign investors than Black startups. But at the end of the day, I guess investors are the ones who make the decisions. So we need to communicate to them that there’s this whole ecosystem of young, smart Africans trying to start companies and solving problems in Africa and, and just invest in them. I mean, of course, you need to make sure you have your decision making in place, but you need to go to the big networks or things like that because I don’t think that most investors are outright racist and they say like, I only invest in white startups but not black startups, No, I think that’s not how it works. I believe in the first place, they need to meet the black from the startups, and they need the networks to do this because most of these introductions happen via personal introductions. Like somebody is connecting somebody with email and or other channels. By then, they already know each other, And there might be a disconnect between some of the local startups and investors in the West. The second half of this problem is that there’s a lot of money investment money in Kenya. Still, it’s mostly invested in things like real estate and classical stuff rather than startups or even AI startups. So I’m trying to get some of the local investors to become angels or VCs. That’s another thing, but of course, the difficulty there is that the investment works pretty differently. If you’re investing in real estate or an AI startup, the risk function is completely different. The market is completely different. So just because you’re a great real estate investor doesn’t mean that you will be a successful AI startup investor. So there are also some initiatives running, some local angle networks founded and that’s great. I might also add that in other African countries, it’s very different. So, for example, Nigeria, there’s a lot of black founded Nigerian native startups that are getting a lot of investment. It’s the vast majority of Nigeria’s local black-based startups rather than white-based startups and, and I don’t understand why Nigeria is like that and Kenya’s like this.

Job description from the Audience who attended the webinar;

Where do you see yourself in the next five years? What’s the best thing that could happen to you?

The best thing that could happen to us is the developments that we are doing right now are working, and we are making this product very successful. But of course, yeah, the path there is not easy. So one thing that we have been working on since 2018 is also developing a blockchain component of our protocol and that’s a pretty important bit for a couple of things. One is that We would like to provide trusted information to people, for the right thing at the right time, but we don’t want people to have to trust us. It’s ourselves a company too much in a way you have to trust us when you’re using the platform, but if you can reduce the amount of trust, then that’ll be great. So for example, if we tell you that your friend likes this auto mechanic, then we could lie, right? That information might be correct, or it might not. And if we just closed the platform, then the only thing that you can do is call your friend and ask them if they recommend this auto mechanic. But of course, the idea was in the first place, that you don’t have to call your friend, that you get this information seamlessly and you don’t need to go to that extra trouble. So putting endorsements like that, on the blockchain on the public blockchain enables people basically to verify the correctness of what we are showing themselves and so that’s how the blockchain idea initially started. But then we added layers on top of that

There is a question in the chat here, for example, about fairness and bias and AI or ethics for using the technology and that’s a crucial topic for us. So, we want to use the blockchain manage paths to enable people to specify themselves what they allow us but also other services on the platform to do with their data. So for example, if people give us enough data so that we can make these context-sensitive decisions, as I explained earlier, obviously, we need a lot of data for that, do we need to know sort of purpose, people’s preferences and what they have and who they know and all that. That’s a tall order to ask people just to give us this data. So we built this platform to allow people instead to exactly say, okay, you can use this data for this purpose, but do not show it to this group of persons or things like this. And then they will be rewarded for that. And so you can decide what data you show and you can also decide ultimately what it would cost users such as our platform, but also maybe other services on their platform to use that data, and then also it can be retracted at any time. Of course, once you have sent some data piece of data to somebody, then it’s out there. And then you have to trust them not to misuse their data. If you would allow UTU to use the fact that you own a Toyota car, only to show Toyota related recommendations to your friends, but you do not allow us to send this information to advertisers. And the next day you are swamped by the spam email from Toyota. So you need to trust us to not send on this data. And of course, that’s a problem that cannot genuinely be solved. At least not with technology that is readily available right now. So there are some things that are more difficult than others. But there’s a mechanism of implementing the way for people to specify what can be done with the data. That’s something we are building and there’s a lot of other problems around this, like how do you make the UI about this understandable and easy to use, because if you give a lot of flexibility around this to people, then that also means that it’s a complex root system that you have that is maybe not easily understood by everybody. So you need to invest a lot of research into UX and UI and how you make this work so that it can be used by not only technical users but by everybody.

Then Bias in AI is another big problem, particularly with machine learning. There have always been discussions around this for years. But the latest development GPT 3, the language model that was recently released which I think is not openly available to apply for better access. But there’s also stuff like, for example, yesterday, somebody posted on Twitter if you GPT 3 challenge up to three with two Muslims, then it generates all sorts of texts that have to do with terrorism and violence. There’s not a single example that the system makes with just random two people which is pretty scary. And then nobody understands these big models because they’ve been trained with vast amounts of data, I guess the data that was chosen to train GPT 3 was also probably automated in some way. And so nobody knows what’s going on inside that model. So that’s a big problem and you cannot quickly solve that, and in my personal opinion, every system is biased in some way. You just need to choose your bias, consciously. If you only train some algorithms on genuine data on the internet, then I mean, most of the internet is still heavily biased towards Western things, or the Chinese internet, of course, to Chinese words. That’s just the history of how the internet grew. It was developed in the US, and then it grew from there and then for a long time, 90% of websites were from America. And now it’s different. But for example, the amount of African information compared to US information is still tiny. And so if you don’t push your algorithms to accurately look at African sources, for example, then you would just get a heavily biased model out of it because that’s what you put in. Then some people are saying, Yeah, it’s just about the data you’re feeding and not about the algorithm itself, which I’m not so sure. But the thing is that with machine learning it’s kind of difficult to really figure these things out. There’s also something where classical AI methods like logic can help better in my opinion, because you understand much better, for example, you know much better. The other rules are how the system reasons and therefore, so what problem with biases you might have and how you might counter it.

Scroll to Top