Dhaval:
This founder built a whole new category of ai, product known as generative learning to help you tailor your learning based on the individual context, learning goals, and learning style. Calin Drimbau the founder of Broadn. Shares how he built his AI product using. Three layers of abstraction and have, you can follow the same type of product architecture to solve your specific workflow using AI resources that he shares.
Welcome, Glenn. Tell us about yourself and your product.
Calin Drimbau:
Hi, pleasure to be here. I’m Calin. I’m the founder of Broadn. We’re building a new solution for learning. It’s personalized learning at its best. in many ways we’re defining a new category and that is generative learning. We’re using generative models to be able to tailor learning based on the individual context, learning goals, and learning style. Of users. we’re very excited to be building in this space and it’s a pleasure to be here and have this conversation with you.
Dhaval:
Wow, that’s like personalized learning at scale. Tell us a little bit about where you are in your product journey. Has it launched? Is it, being billed? Is it in beta? Is it still being developed?
Calin Drimbau:
Sure. So we’ve been on this journey for about a year now, and we’ve launched a couple of products., the first product that we launched was a product that was doing classification on, podcast content. So we would be listening and, transcribing the text from audio and then identifying topics that are being discussed in conversation and using AI and then clipping creating automatic clippings and placing all of these into a platform for learning which was in a form of a mobile app. So that was our first product, and moving away from that, in conversations with , our users, we’ve learned that what they wanted to do more above and beyond getting the best clips from podcasts. They wanted to navigate and explore this content by searching. so a big problem for people in the audio space is, identifying the most valuable parts of a conversation. , and they wanted to do that by search. so our newest product that we’ve recently launched is, A semantic search engine on top of podcast content. and happy to speak more about that. We’ve launched that last week on Product hunt.
Dhaval:
Wow. Yeah. Tell us a little more about how did you find , the kernel of the idea and, how is it doing now that you’ve launched it? How is it received by the audience.
Calin Drimbau:
Sure. So I’m a big podcast listener, and I’m a big consumer of knowledge, if you want from books, articles, YouTube. I consume a lot of information and my personal problem was that especially with audio and video content, it’s not easy to navigate, this content. It’s usual. It’s usually presented in a linear format. So oftentimes , when I’m looking. To consume content is because I’m trying to solve a problem or it’s been, it’s because I’m trying to learn more about a mental model. So having had experience building machine learning products on text, I thought, why doesn’t anyone do? Processing and parsing of podcast transcripts to identify and classify what’s being discussed. So that was the genesis of the idea. Beyond that, I suppose once we’ve launched it , in the market as I’ve said previously, users wanted more flexibility on how they consume this knowledge the app itself was very useful in terms of like saving a lot of time and instead of having to listen to one hour or two hour podcast, they’d be able to listen to 10 clips on the specific topics that they wanted to learn more about. A lot of the content on the platform is learning content, like entrepreneurship or product or any type of lessons and clips from , podcast surfacing and talking about these topics. But then one of the constant requests that we’re getting from users is I wanna have more power in my hands to be able to navigate this content. So if we take Lenny Rachitsky’s content, for example which I’m a big listener, of content too, they asked for an ability to search for specific topics or episodes or experts so that instead of listening to, the two hour episode, they’d be able to zoom in and double click on exactly the specific lesson that, that the author or the guest is, highlighting in, in that episode. So that’s why we built the search interface on top of podcast content. And we’ve actually built it. Just on top of Lenny’s content as a first drop, obviously the same technology can be deployed and adapted to surface semantic search on any podcast content. And this is one of the things that we are, we’re exploring but, podcast content and searching, or semantic search, if you want, is just in our view, the first or the stepping stone in term in terms of realizing our, our vision for personalized learning.
Dhaval:
Wow. Yeah. I would love to get into the nitty-gritty of how you build the product. We’ll get into that in a second. But first I wanted to dive into the user journey. So you have two-sided market. One is the listeners and the other are the creators. Are there any other sides to your market? How do you monetize this product? Is it subscription based product? Is it advertising? Yeah. Tell us a little bit about any other sides of the market you may have, and more importantly, I’m curious about is this like a one central place where the end consumers go to search for podcast clips? Or is it a service you provide to podcast creators who want to put this interface on top of their podcast so that people can consume it more effectively?
Calin Drimbau:
Yeah, good question. Obviously these are things and ideas that we’ve been exploring for a while now, whether we pivot into one side, a two-sided marketplace, or a one-sided marketplace for , for creators. was the questions that we’ve, a question that we’ve asked ourselves At the moment. Our vision is to create a consumer platform, not a typical marketplace if you want, because we’re taking all the public content that is exists out there. So everything that a creator has deployed and it’s not just audio. We started with audio and we started with podcast. But we’re able to process video, YouTube and we’re able to process essays and any type of blog published by, by authors and creators. in terms of the, monetization, cuz you’ve touched on that, the big goal is for all of these mini products that we’re building and validating at the moment to form. A bigger consumer product that will have a multiple set of use cases. So indeed one of the first use cases that the product will feature is the search, or semantic search, throughout a set of podcasts. But above and beyond that, the additional use cases that we’re currently working on building in, into the platform are the ability for a user to define and set their learning goals. , and, the platform, it would be able to know and understand the context of the user, and it’ll generate, if you want , a dynamic personalized course. For the user. Now, the way it does that is, is a little bit dissimilar to, a typical interaction with something like chatGPT where. You’d be able to prompt the interface and ask for content. And a large language model will process that request and give back some answers to you. We’re doing it slightly differently in the sense that, we have micro verse of content that we are pre-selecting, and on top of which we deploy semantic search. So this is why the semantic search, is very important because within the universe of content that is created by. We’re first running semantic search, and then we’re deploying, summarization and other typical, generative AI techniques to surface that content for users and, and match, their learning goals, their learning style, and the context for which they’re asking that. So that will be the main product that we’re building. That product is currently being built. But it will take the form of consumer subscription model to answer your question on monetization as well. Yeah.
Dhaval:
You touched on overall conceptual flow of your product. Let’s, let’s, dive a little bit in there. So you mentioned microverse. And then you mentioned the ability for semantic search to be sitting on top of that. So if you are to help us understand the conceptual product architecture for the people who are product creators, the audience of this show is product creators, are product creators who either want to build an AI product. And don’t have the deep AI expertise or people who want to infuse AI in their existing product. So we try to unpack a little bit about how the product itself is created here not too in depth, but just a high level overview, conceptual overview. So you mentioned that, for a user’s. so for your product for a user’s learning goal. You would establish a microverse of content and on top of that. You would have semantic search and then on top of that you would have generative ai. Did I get that right? Please correct me on the the stack and the overall flow of information so that you can help users the end users that you have achieve their goal what is that product stack? What is that architectural data flow looking like?
Calin Drimbau:
Yeah, that’s actually pretty accurate. So there’s three layers if you want, and a pipeline of information flowing from this content, monitoring and extraction engine. So this is what I previously called our. Universe of content our microverse of content because we’re not mining the whole internet. We’re starting, for example, with entrepreneurial learning, and in this case the definition of that micro verse of content would be a pre-selected list of YouTube channels, podcasts, and articles, blogs or creator sources that will form all the data flow that, enters the the universe of content if you want. So there’s a the first piece is basically a content ingestion piece. It’s not a sophisticated piece in here. We’re not talking about ai. It’s purely subscribing and listening to a set of content. And whenever new content gets updated on these channels, we parse that information again in the second stage of the process. So now the second stage of the process, or if we’re looking at it as an infrastructure layer, the second level or the second layer. Is the layer where we’re deploying search. So on this layer our approach was to use the form of semantic search. We’ve actually built, our, and adapted our own algorithms in here. But for anyone that is new to for example, building or defining their own vector embeddings algorithms, there’s already a set of pre-existing algorithms that you can use. You can either go down the route of using large language models. Open ai, the classic example, provides an API and you can use their embedded endpoint to do semantics search. Or you can go to Other providers like Algolia that have been provided providing these type of services for a long time or, pick up some of the smaller transformer models that exist on hugging face and adapt them in a way that makes sense for your particular use case as I’ve said, we’ve went down the route of picking some existing smaller transformer models and adapting them and combining them in a way that makes sense for our own use case. So that’s the second part of the process, or the sec second layer of the architecture, depending how we look at it. The last layer or the final processing component is what we do with the relevant content that, we find through semantic search. And in here again, you’ve got multiple options. So in order to get key ideas summarized, you would, again, need to use some form of intelligent processing, right? So the simplest example again is using an endpoint from Open AI using their API and basically feeding as context, the results of the search and getting that, endpoint to summarize your results. Of course, open AI is not the only solution. They’re multiple, large language models like Anthropic out there doing similar things. Or you can go down the route of a small language model as well. So all of this really depends on. Your level of familiarity with ai whether, if you’re a non-technical co-founder, there’s still a lot, or you don’t have a technical co-founder, there’s still a lot that you can do an experiment with without having, deep AI expertise because a lot of these interactions with AI models these days are in the form of a simple API query requests, and ultimately whether you need the complex AI function to be built in house or whether you just use existing models will depend on. The type of market that you’re addressing. Whether it’s business or consumer, what type of industry you are in, and whether you’re building a small SaaS project or a venture backed business. Everything’s possible this day and these days. And then choosing the right solution, if you want, really depends on on your goals.
Dhaval:
Wow. You really summarized it Well, thank you so much. This is gonna become a masterclass on how to build a semantic search engine. And feed your users small pieces of highly valuable and summarized pieces of content. I would love to, I would love to unpack, one more piece of information, which is do you have a very technical background or does your team, do you have a deep technical team, or tell us a little bit about the context and the expertise of your team.
Calin Drimbau:
Yes. So, I have a technical background. I have a computer science degree and , I’ve built products in, in AI and using natural language processing techniques, for six to seven years. So I personally have some experience. Building this, but I’m not technical in the sense of coding. So I don’t code anymore. My expertise or deep expertise is in product. So doing anything from product discovery to execution, that’s where most of my career has been focused on. My co-founder, however who is complimenting me on my skillset, has a deep. Machine learning, expertise. He has master’s degree in machine learning and, also has spent 16, years or more building machine learning solutions.
Dhaval: How did you convince him to join you?
Calin Drimbau:
it was, one of those fortunate situations where we worked together, actually in in legal tech, a legal tech company building together, another AI powered solution. So we worked together for a year and a half. , and I think we’ve realized that we work together really well, and this is a big opportunity to use the new wave of technological advancements to redefine an industry, to redefine how we learn. We’re both excited about lifelong learning and we both have the right level of expertise to build this product to live. So it was almost a no-brainer for us to work together
Dhaval:
wonderful. And tell us a little bit about your financial journey. Did you bootstrap this company? Have you raised capital? Have you raised a round of seed? Are you looking to raise it in the future? Yeah.
Calin Drimbau:
Yeah, so we’ve been bootstrapped so far. we’re lucky because between him and me we are pretty much covering all aspects of the early stages. So I do product and design and he does development. I also do the marketing side of things with a little bit of external help. So between us, everything that we’ve done so far Wasn’t that expensive to build if you want. It was just more our time. and I think because, we wanted to move fast in many ways, it was better for us to get our hands dirty , and start building. That’s. Probably gonna change soon because given the, feedback we’ve gotten on the recent release and the previous release and the fact that we’ve built a big wait list for our products, we’ve received a lot of, interest from investors. So it’s probably the right moment, and opportunity in time for us to raise the capital to pre seed round to accelerate the , the growth towards, the vision of, personalized learning.
Dhaval:
So One of the primary data sources for me to consume audio information is Audible, which is the Amazon Audio Bookstore. Is that, how do you get access to audiobooks in your data, processing? Is that part of your roadmap? I’m just curious.
Calin Drimbau:
It’s a good question. It is something that we’re considering, for the long term. It’s not something that is, immediate in our roadmap if you want because with books, there’s a little bit of complexity with regards to copyright whilst podcasts are out there in the open public space, if you want, and we can process that content easily by just respecting the copyright and playing back content giving the podcaster is more reach if you want. The situation is a little bit more complex when it comes to books, and you’ve probably seen that in the world of ai there’s a set of lawsuits that are starting to happen left and center with regards. To copyright, both at the stage of, data used to train the models, but also at stages of how the content is being produced and how the data is used for, especially if the data is used for, financial gaining purposes or profit making purposes. So at the moment, it’s something that we have in the back of our mind, but we haven’t looked or researched into detail what sort of rights we need to have in place for us to be able to process and integrate books in our solution.
Dhaval:
Great. And with that we’ll wrap up the show, but just one last question. You touched on something, you touched on a very important topic here, which is lawsuits and the source of data and privacy and copyrights. Right. So AI product creators need to be very mindful of the source of their data, the privacy and the copyright details of that data before they try to profit from it. What other advice do you have for AI product creators who are aspiring to build successful products in this space?
Calin Drimbau:
The most important thing is not to rush. , I think there’s a certain frenzy in the market now given the abilities that the technology offers. So everybody’s rushing in to, to build cool products. But in reality, for any founder, I think the important question is what is the quickest? What is the quickest route to value? So beyond and outside of building a cool in the product tool, what is it that you need to do to create value, either for a consumer or for business? And start from first principles of speaking to your customers, understanding a problem. Not start from the hype and not start necessarily from the technology itself. But once you figure out what problem you’re tackling or how you want to use the new technology, then it’s the right moment in time for you to accelerate. And, the APIs that are out there from large language model providers enable you to quickly experiment like never before. You can build solutions and products that used to take months, you can build them in days. So once you know what your direction is and what sort of business you wanna build, Iterate quickly, launch quickly, see what works. And then once you’ve landed on something that attracts the, user’s attention and provides value to them then scale.
Dhaval:
Well, thank you so much, Calin. We really appreciate you joining this show, taking time to share your hard-earned knowledge and looking forward to have you back on the show in the future when you have more or lessons to share with us. Good luck with your pre-seed round. I’m gonna be following your journey and looking forward to connecting with you here in the near future soon.
Calin Drimbau:
Absolutely. Thank you so much for having me. It’s been a real pleasure chatting to you thank you.
