Ep#94: Demystifying Machine Learning with Michael Louis (Founder, Cerebrium)

[00:00:03] Adil Saleh: Hey, greetings everybody. This is Adil from Hyperengage podcast. And, you know, time after time we get to meet these technical co founders. We get to meet products that are extremely technical and that are dedicated more towards technical teams, developer teams, and we get to learn a lot from how everything at the back end of a product goes as a subscription. Based product or business, how the real back end experience looked like and what kind of heroes we have at the back end, such as the CTOs developers, machine learning experts and all the crew of technical team that brings the best software experiences out for the customers. So we today came across Michael, who's the co founder of Cerebrium. It's a SaaS platform that recently got accepted in the Y Combinator accelerator. That is quite a success. And they are doing it at scale for machine learning experts, machine learning teams to deploy large language models at scale and with ease. One of their competitors quite reputative in the market, that is Amazon, they have their own cloud service and they're absolutely providing a lot of big enterprise platform. They're hosting their machine learning models on their cloud. So we're going to be talking more about Cerebrium today and specifically this gentleman, Michael. Thank you very much, Michael, for taking the time. [00:01:41] Michael Louis: No, thanks so much for having me excited for the session. [00:01:45] Adil Saleh: Love that. So, looking at your background, the notes I got from the team mean you've been a CTO of a company that's named one part that later got acquired by one of the Walmart entities and you spend most of your time when it comes to delivering products in South African region. So could you tell us a touch more about what is different about that religion? Because we had a couple of platforms such as Warmats, that's fleet management platform out of South Africa. We had fintech came up last year from South Africa too. And that's kind of a booming market when it comes to fintech when it comes to core enterprise tech. So could you touch us deeper into how does the South African market treated you throughout this time and how you guys connect with it? [00:02:36] Michael Louis: So, I mean, our entire founding team is and even our entire team today is 98% South African. And South Africa is a very phenomenal place and very interesting place because on the one side we've got a lot of political instability, lack of funding, a lack of technical advancement and innovation. However, there's also just so much technical innovation and advancement even in our fintech sector, like you mentioned. So starting businesses there is extremely challenging. We have sometimes 8 hours, no power a day. So how do you even get internet and power for your computer? So, a lot of kind of challenging conditions. However, I guess it's made South African entrepreneurs a lot more kind of persistent in their pursuit to succeed when building a company. Other things that are pretty different is the South African market is pretty small when compared to European or American counterparts. For example, South Africa has got a population of 55 million, but I think less than 7 million are taxpayers just because of we've got extreme inequality and so a lot of kind of different adverse effects. But I think it makes us only strongest as entrepreneurs. [00:03:44] Adil Saleh: Very interesting. And that's why that's the sense of entrepreneurship is where you see yourself making the most impact in industries like geographics and spaces like in South Africa, South America, that we get to speak a lot of startups in Argentina and Brazil. So now, talking about prospective AI, I know it's kind of not a very new term out in North America, but how do you guys create awareness and customer education while thinking of building a platform like this? And then we'll talk about your initial journey and how it all led to combinator and everything. [00:04:29] Michael Louis: Yeah, so I mean, to kind of tell you how we came across the idea of Cerebrium is it was from our previous company, One Cart, which was an on demand grocery delivery company. And one thing you'll know about kind of ecommerce is that the margins are typically very low. So I think about 20% roughly. And we were growing massively during COVID and even kind of before that where we grew from about $150,000 actually even less a month in GMV to doing more than $200,000 a day in seven months. And one thing that kind of made us want to improve on our margins is using machine learning to do rooftop optimization, inventory forecasting, cross selling and upselling products. And so when we kind of as a team try to tackle that, it was probably just the worst experience I've had implementing a software product in my life. Tooling was fragmented, it was expensive, no one really knew what best practice was. And I think the biggest kind of concern, which we're trying to solve a stream, or at least one of them, was just the risk for a startup or even like a Series A or scale up to take when embarking on machine learning. And the reason for that risk is you have to hire a lot of upfront resources and costs for something that doesn't have guaranteed success. There's a famous stat that 80% of models don't make it to production. And we obviously have seen that through and through where you invest six months of time and resources and models don't perform well in production. However, the models that did make it into production on One Cart completely changed our business. And we got acquired by Walmart about two years ago. And so me and my co founder, who was the lead engineer for our company, we kind of know what if we could make machine learning more accessible to smaller medium enterprises in order to grow their company to great heights and change the way that they fundamentally operate as a business. And so that was kind of the initial kind of thesis that we wanted to go out and solve. And we've kind of gone through a few different routes to try solve there. But as of the last, I guess, kind of 18 months, we belong to Rebrim. And it's been a pretty phenomenal journey. [00:06:40] Adil Saleh: Absolutely. And that's very interesting to first make sure you live with this problem long enough and you feel the pain and, you know, then being somebody on the other side, you got to make sure that how to solve it. And for an engineer, engineers, I call them as problem solvers, it doesn't have so much to do with the degree and they try to find ways to solve the problem. The biggest thing is that you got to make sure they know the problem inside out. They have been on the other side of the table and that's what you've pretty much done. So what kind of synergy you guys have? Of course, you discussed with your co founder at that point and what was your thought process into making it a monetizable model and a model that can have a SaaS experience because in these kind of product, it's so hard to create a SaaS experience that is kind of self served. So what kind of efforts that you guys drove at that point? In the beginning. [00:07:42] Michael Louis: Myself and my co founder have been kind of engineers our entire life. And I think everyone always says engineers are very different. We don't respond to cold emails. We hate to pay upfront for things without testing it. We want to know the details, we want control, we don't want vendor lock in. And so those are kind of key considerations that we try to kind of set from the get go of our product. And a couple of kind of key pillars that we've gone is. Abstraction is always a very fine line as a developer product to kind of tread because abstract too much and you're kind of locking them in and preventing them from doing stuff they want to do and abstract too little, and then it's kind of like, why would I not use just kind of the base frameworks to actually build a product? So abstraction has always been a big kind of content in our team and we always discuss it. Two is developers have to be able to try a product to kind of its extreme before even implementing something. Production three is customer support is probably one of and documentation is one of the most important things for any developer product. Reasons being you don't want them to get any blockages. And if they have a problem, how quickly can you unblock them? Because if a developer gets blocked, you've pretty much lost them because then they start looking for competitors or alternatives to develop. Because I guess developers always under the gun in any company when it comes to Sprints that they have to meet. And so those are kind of in three things that we've always had on so customer support and documentation, self serve, and kind of the right level of abstraction. [00:09:05] Adil Saleh: Definitely we will dig into deep into how you're basically delivering customer education because we spoke to Andrew from Propelaut, the guy Robin, I guess from Nango. Both of these are developer tools, one for authentication, the other for API integration and all. So and it is super important to make sure you get all of your customers in a unified space where they can share knowledge, where they can share problems, where they can share use cases. And it's sort of in a community led growth that you're trying to approach as a business. So, looking at your approach, joining Y Combinator last winter, what was the biggest reason for us? [00:09:48] Michael Louis: The innovation of machine learning was in the US. And we were South African co founders and we wanted to get over to the US to, one, learn from the best to hire, I guess from the best pool. I mean, we also think South Africa's got a phenomenal pool, but unfortunately, in machine learning, that's not really the truest case. Three was for the network, both from like a business standpoint, from an investor standpoint, because even after we sold our company to Walmart to raise money in South Africa was still difficult. And I guess kind of the fourth reason maybe it's a bit egotistical is we also wanted to compete against the best. We kind of believe that founders push other founders to operate at a level that you wouldn't normally compete at. And me and my co founder like competition and we've been meeting and competing in some of the most incredible founders. [00:10:36] Adil Saleh: I love that. I love the fact that you're trying to have a peer group, amazing peer group where you guys find people hustling. They are grinding every single day to make an impact. And that's the biggest reason that's why they're successful, because they are getting look like founders, founding team, I would always say, and I'm open for debate, that why commoner people, they are more investing or betting on people than the technology as compared. So they always want people to be bigger than the technology in the first. So because it's people that drive technology, their grit, their perseverance, their passion and all of that, they talk so much about the co founding team and their relationship. So how was your relationship about that? And just give us a sneak peek into your introduction into the YC team. I'm sure you might be might gotten interviewed by Michael or one of his team. So just drive us through the experience. We ask this a lot to all. [00:11:37] Michael Louis: No, no, it was a phenomenal experience. It's so funny. I think the one thing about YC is you think you're going to go in there and get the golden nugget of how to become successful and their advice is just so simple that you actually start laughing to yourself when they actually tell you advice. And obviously they're true. I mean, they've been doing this for decades and I think you kind of said it right. The one thing that YC's done, and it's maybe not spoken enough about in conventional startup wisdom is, and I tell my team every day is pick your team before you pick an idea. Because one thing that YC will even tell you is whatever you come into YC in with as an idea, you're probably going to leave YC with a completely different idea or just reimagine the way that you've done it. And so definitely pick team first before picking an know these are the people that you're going to work with every day. And I can tell you, even if your company doesn't reach the heights that you wanted to, you'll never regret that time with them. And then yeah, the advice we got from YC has been phenomenal. Just the community of founders that you get access to, it's been pretty phenomenal. [00:12:40] Adil Saleh: Yeah, and the interesting fact is not just you get a piece of gateway a network eventually customers. I know a lot of you guys understand your technology and you share knowledge and products and you try your products. A lot of your initial customer base must or might have come from YC network, which is really good. The interesting fact is the way they are driving strategically, like they are pairing you guys with the best advisors, best researchers, best technical resources, best panel of investors. There is so much that is on top of just getting you see a lot of accelerators that are doing the mainstream coaching and growth and all of that. [00:13:29] Adil Saleh: So in terms of strategic partners up till now, it's been almost a year now. It's going to be a year now. So what did you get as an outcome when it comes to strategic partner with an accelerator such as YC? [00:13:43] Michael Louis: Well, I mean we got obviously investment from both them and just kind of the YC community that they invested through. Whether it was kind of venture funds, whether know angel investors. We have YC companies that are some of our customers. We obviously use a lot of YC tooling as just, you know, I don't even know if you know but still today we can meet, know the YC partners like Mike or Dalton or Tom and actually ask some advice. So they are on your cap table and they're constantly available, which is great. And just the YC internal, they have an internal website. It's just full of resources about what people have done to go viral on hack and know what are some tactics or metrics that you should reach for series A. So yeah, just a phenomenal resource to have all. [00:14:28] Adil Saleh: Amazing, amazing. So let's talk about a little bit about Cerebrium your product like you started about two years back about two years back. What is your go to market post? Let's say handing over validating the idea you have customers, you're doing the community of getting them all in one place. What is that one thing that you think, okay, this is something that we need to do five years later as well. This is something that we nailed it. Then we just need to scale it in terms of go to market, like in terms of market positioning, in terms, of course, growth in sales and valuation. [00:15:09] Michael Louis: I think one thing with go to market is it's always very interesting because I think it's very tough to say every business is the same. This is my fourth business, my first B to B one, and it's a completely different go to market strategy as previously. So I definitely feel like I'm starting a company for the first time again. But in this one, I think a developer product is always best well known as a product growth strategy, just because developers trust the opinion of other developers. And so an outbound sales motion doesn't really work because investors don't take meetings. Sorry, developers don't take meetings. You can write as much content as you want, and that's pretty good, because if you can do tutorials of how developers can implement things, then they get an idea of the capabilities of your platform. Google ads don't really work. I would say partnerships, I would say, is a phenomenal one. How can you partner with other tooling on opposite ends of the value chain of your product? So, for example, with us, we've partnered with some SDK companies. We've partnered with monitoring companies, observability companies, just because it sits on kind of either side. And so those have kind of been our way of tapping into communities that we obviously don't have access to. And that's been pretty phenomenal for us as well. [00:16:21] Adil Saleh: Yeah, absolutely. And partnering up with technology or tech stack that complements yours, and it's a win win for both businesses. From a business standpoint, let's say we came across a platform that is more on the marketing automation side, and then they basically collaborate with a content automation or AI powered content. So this basically creates a synergy. So you're looking for some partnerships that are healthy for your business use cases and they somehow complement longer term. [00:16:53] Michael Louis: Yeah, I mean, I think that's the lesson I learned from One card is instead of onboarding individual retailers, we would go to retail chains and we would get 200 stores over a couple of meetings. The same thing. Now we're going live with another partnership kind of in a month's time. And that's with a video SDK company that they can integrate machine learning into video SDKs. And with that, immediately, thousands of their customers will get access to machine learning capabilities right inside this SDK. It's one partnership. It's thousands of potential customers. Even if we close 1%, it'll be pretty good. [00:17:31] Adil Saleh: Great. So now talking about your pain customers. How big is your customer base? Let's talk about it's. A developer to active customer base. What number is it at right now? And what are your growth metrics going forward? Maybe year to year or quarter to quarter, however you guys go. [00:17:46] Michael Louis: Yeah, so we go kind of month over month. So at the moment we've got hundreds of customers. Obviously paying customers is kind of the pareto principle. Only a portion of that is paying at least significant amounts. I mean, we have tons of hobbyists who are paying $10 a month, but we don't really count that. But hundreds of customers, we're growing about 30% month over month sometimes. What's quite interesting about the AI space at the moment is we've seen our growth double in a week because some AI app went viral, but then three weeks later it's dead. And so that's been pretty interesting to keep up with. But yeah, that's kind of where we're going up now. And we think with a lot of stuff we're releasing over the next two months, we think that's going to drastically increase. [00:18:28] Adil Saleh: Love that. So we actually get to hear this a lot in today's day and age factor is going on. So there was just talk about retention, business retention, revenue retention, more on the customer success side. So in this day and age, it's harder to retain the existing install base than acquiring a new one. Do you agree with this? [00:18:49] Michael Louis: Yeah, so, I mean, they always say it's better to upsell existing customers than try and buy new ones because your CAC is obviously a lot lower. We've seen our retention is above 52%, which is great. Obviously we have plans to hopefully make it higher. We definitely have had churn. However, churn has mainly been the reason. Maybe it's economic times, maybe it's the cost of GPUs, but churn has mainly been to cost reasons. Some of our users have been given free credits by large cloud providers. They've been given free GPU access by investors of theirs, angel of theirs that they end up churning. But yeah, hopefully with the releases over the next two, three months, we'll hopefully address a lot of those issues. [00:19:30] Adil Saleh: We are building a B, two B SaaS that's sitting in the beta and we got charged around 700, $800 anonymously by Amazon. I'm sure we had our credit card in it and everything, but we never knew it's coming because the credits run short and we never knew it. And now we're trying to move the server to save and optimize the cost for a bigger user base once we're going to have all the users. So I'm not a technical guy, but my CTO tells me all of this. I don't have a technical caliber to talk to you about all of these servers and cloud based deployments and everything. [00:20:04] Adil Saleh: So now thinking about customer success, how you guys are ensuring success post sales. Just talk about paying customers, how you're making sure you're sitting on top of their product usage, platform usage, how you're setting metrics to ensure the adoption of the platform. Actively, of course, retention of those features that you have, some modules that you have. So what kind of process you have around that, any tech stack you're using around those to ensure these parameters. [00:20:37] Michael Louis: I would say it's probably a lot more simple than it should be. The biggest thing that we track is basically our metric is how many inference calls are people making on our platform and is their business growing? And we obviously don't track how active a user is because sometimes they deploy a model and they're working locally and they might not come and deploy on our platform. So what we do is we track, is their business growing and every month we have an email that goes out to all of our customers that reaches out saying, what are your biggest issues right now? Is it even related to something that we can solve? And why is it an issue? And I think the whole reason for that is trying to figure out what is the next thing our customer might face as an issue. Not even right now, because if we solve them, obviously they might retain longer, but as well as new customers might see that we have that feature as well and they might be further along than the existing customer, like their journey that we might have this feature. And so that's kind of a key thing. It's just remaining extremely close to your customers. And in terms of kind of outbound or at least kind of reaching out as top of funnel is always very important. So it's always very tough to do with a product led growth because you don't really know where your users are coming. We know they're from blog articles, we track, we can see they're from referrals, we can see that they're from partnerships, they're from influencers. But other than that, it really depends on the business use case. A business of sometimes two people can be just as big as a business of 20 people in our space at least. And so it's very interesting to actually see. [00:22:09] Adil Saleh: So a lot of your organic, I would say customer base is coming from content, right? You guys have a blog. I do see that. And I'm sure one of your, you or your co founder are extensively writing those blogs for knowledge and problem solving for these machine learning engineers. And that's how you basically search their problems and that's how you get ranked. So on the search engine marketing, are you doing organic standpoint? Are you guys doing anything other than writing blogs? Like value added blogs? [00:22:44] Michael Louis: Yeah, we do like partnerships, like I said, with companies. We obviously do a bit of paid ads, but it doesn't really give the return. It's more just something we're testing at the moment. And then, like I said, word of mouth and then also just the content is very interesting because I don't get to code as much as I used to, which is quite sad. So I told the team I'll write all the contents because it gives me a time to write some code again and kind of publish it. [00:23:08] Adil Saleh: Yeah, it becomes like sometimes you as a leader have to figure out what is your high value task. If you can delegate some of the tasks that are getting done as good, you can diversify yourself all these technical tools. There's no better writer than a technical co founder himself or herself. It's so powerful. There are some platforms I don't exactly recall. There are portals where they have peers, where you can rank some of the problems that you're saying. That's how you get traffic. It was some platform only for developers. It was where they can present their code, where they can add their portfolio of code. What's the name of that platform? [00:23:53] Michael Louis: Yeah, I mean, the one thing that's quite interesting is we've seen ourselves that we've tried to actually release content on doing things of like how does stable diffusion image to image work? And to be honest, the content just doesn't work well with our audience because they actually want to see how quick is the latency, how good can you actually get these images? What are the prompts that work? How can I swap out maybe some of the encoders that are using the model? And so we've seen that our audience really likes kind of the deeper understanding and that's something that you can't unfortunately use GBT Four for. And you can't use just any content writer for either. [00:24:29] Adil Saleh: Yes, exactly. Because no matter you can spend an entire day prompting the GPT Four, you're not going to get the relevant answers or the answers that you deserve because it's more technical. Or else you can build your own prompting tool to address these kind of issues on top of GPT data layer, overlap your data points and it's like machine learning. You can create your own model of machine learning. And for that, again, you will have to use a platform like Cerebrium. So now, thinking about customer education, I know that there's so much of product stickiness. People that come in deploy their resources using the platform, they tend to stick for longer if they don't have any unprecedented problems. So how you're ensuring the education, ongoing education of the customer, that's regarding the use cases, regarding the product, platform usage, all the navigation of the customer education. So what kind of platform is that? The blog. [00:25:29] Michael Louis: I would say it's our blog. And then just actually through our communities and then also just, I guess our actual products itself. So I mean, the one thing that we say with kind of the space that we're in that's completely different is our consumers, while they're also our customers are also engineers, but they're also machine learning engineers. And one thing is a lot of engineers don't know the specifics of machine learning and they've really had to scale up the last couple of months. And so we've had to create abstractions to such an extent of how do you kind of convert these 21 million software developers into machine learning engineers that are capable of releasing these products? And so we've kind of done two things. One is abstract the infrastructure. Don't worry about that. Like the GPUs and the CPUs and the dependencies and scaling. Just focus on the use case because that's obviously, at the end of the day, the most important thing. And then the second thing kind of what you mentioned is the research of machine learning. So machine learning, whether it's like computer vision, whether it's regression, which is traditional machine learning, or whether it's large language models, they all have their specific nuances, how you train them, how you structure your data, how you influence them. And so we've tried to kind of abstract as much as we can or educate our users on how can you actually implement this yourself or how can we abstract away that. You don't have to think about it unless you're an expert. And so I guess that's with the constant what we spoke about, the customer success, just constant communication with your community, your customers, and then, I guess, constantly investing in use cases yourself as a team, we're constantly testing out the latest stuff that comes out from Langchain or from Meta. And the. [00:27:02] Adil Saleh: This is this becomes very interesting at scale. Like if you are thinking of having large machine learning departments at a company, a very big enterprise tech company, legacy companies that cannot migrate their entire database to a technology that is available in the market right now, they cannot shift everything because they're big enough. And that ship has pretty much sales. So they can have tools like Goto platforms where they can deploy all the machine learning models and basically cloud it, host it at a limited cost. [00:27:39] Michael Louis: Well, I mean, the one thing that's tough with that at the moment is previously everyone's trying to use LLMs like an API, constantly interacting with and engaging with it. The only difference is that typically requests were run on CPUs, which were 1,000,000th of a cent for a second GPUs, depending on which ones you're using, are hundreds or thousands of a cent. So now these large enterprises are looking at an increase, they're the same traffic, but an increase of 10,000 x in their cloud costs. And obviously, they can do long term contracts with these vendors and that's what they're doing, or they can buy their own hardware, but it just shows how inaccessible this is to companies who are truly at scale. But like I said, we have things in kind of two months that we think will make it I mean, we make it pretty accessible now, but in two months, I think there's going to. Be something very cool that the community. [00:28:27] Adil Saleh: Will like, wow, amazing. So shout out to this guy and we'll definitely check back with you in two months. And if you have any content fees that you can write about it while you launch it, so then we can distribute it across our community. As now, you know, we tend to share any resource or anything that comes to our network. You mentioned the platform, the API platform that you're trying to collaborate with. One of them can be Stoplight. One of my friends works as a product manager there. They recently got acquired by one of industry giant I don't recall the name. That's also a technical tool, so I don't recall the name. It's competitive to postman. They're pretty as big as postman. It's a multi billion dollar business. They've acquired Stoplight for one of their product lines. So if you want to collaborate or anything, we can have your introduction with one of their leadership. So Stoplight is something you can check out. And I think one last question is are you planning on hiring? You need anything regarding at this point in time of scaling the team, on the technical side, on the customer facing side, anything that you're trying so you can share it out loud to people listening, what kind of traits you would wish people have before applying to Cerebrium? [00:29:50] Michael Louis: Yeah, I mean, we're always hiring, looking for kind of engineers at this stage. That's who we're kind of looking for. Obviously people that have experience in the space, people who've previously built businesses are always a plus, but at the end of the day, we always say if you have the right attitude and you're willing to learn, we'll definitely spend time investing in you as you grow. And yeah, what we typically look for, I mean, I think everyone always says they hire the best. I think no one ever defines what that is. And I think for every companies it's different. For us, it's someone who is extremely good at what they do. Two, has a willingness to learn, is extremely curious, and to be honest, like reliability. Can I count on you to get a task done by when you say it's done? And Five, which might be a little bit controversial, but someone who challenges every member of the team every day. We love it when someone says, I think that architecture is not going to work, or stupid, and then they tell us why. So, yeah, definitely someone who is confident in their ability and obviously someone that's always great to work with. [00:30:53] Adil Saleh: Very inspiring. Very inspiring. So thank you very much, Michael, for taking the time. You guys reach him out with Michael Reese on LinkedIn. And anybody listening to this can chat regarding anything related to Cerebrium or related to machine learning models that you can deploy for your company, for your own engineering infrastructure. I think this can make a big impact. We are pretty much heads up and pretty much gears up for what he's going to be doing in two months. And we'll keep you guys posted. Thank you very much, Michael, for this time today. And you've been such an insightful and powerful and knowledgeable center for me personally firsthand, and then once it goes out, it will be for the audience. [00:31:38] Michael Louis: Thanks, Adil. Appreciate it. Really enjoyed the session. [00:31:41] Adil Saleh: Love that. Have a good rest of the day. [00:31:43] Michael Louis: Good cheers. You too. Bye.