Tips for Adopting AI and LLMs in Business: Lessons from Michael Vandi

And input, you know exactly the kind of input that was going to be, you know, displayed to you. So if you look at it from the context of lowers, you have, um, let's say a W2 that you need to process. And let's say you put all of the numbers into an Excel spreadsheet and then you did do like a sum and then it gets you there, you know, total income and whatnot. Welcome to the OTECH Talks podcast where it's all about working on technologies. You will hear the latest tech information and your host Cash Eve Manzor brings together product overviews, howtos, best practices, tips and tricks and troubleshooting techniques. Today we have a guest speaker with us, Michael Wendi, the CEO of Adi AI and as the name says, you know, Adi AI. So it must be something with the AI. So we're going to hear from Michael, Michael, welcome to the show. Thank you so much for shape. Happy to be here. So Michael being an open tech talks first time as a guest speaker and the true is nowadays focusing on artificial intelligence, gender, the AI and the objective is to give something back to society, so that based on our experiences, the audience can learn from us. And being you as a you as expert today as a speaker. So if someone a new listener, how he will understand who is early, who's Michael, what he's doing and what is this adi AI. Right. So a little bit about me have always been interested in software engineering computer science and tech. I learned how to code when I was 14 growing up in free town, Sierra Leone and then I moved to the United States for college undergrad study computer science and then worked at AWS for a little bit. I also did a bunch of extracurriculars and in college and during my time at AWS, I just quit AWS to do my masters at CMU, focusing on software engineering, I did some AI classes there as well. That's one I deeply became interested in artificial intelligence and started tinkering with AI and building on top of large language models and eventually build a tool for myself on YouTube life that was adi. It was never meant to be a company. I just built it on YouTube live and then people started requesting access to it and then it became a. So and it's kind of form adi AI what we're doing is we're building the world's first AI loan processor. Okay. So right now it takes anywhere from like 45 to 60 days, even higher to close a mortgage loan. And that's because the loan officers have to do a lot of manual tasks like reviewing documents, managing client follow ups, doing due diligence on their borrowers. So we put all of that into AI agents and then help these loan officers to do these documents and then close loans up to 90% faster. Okay. Now, as you said, you you just build up on a video and people start asking it, but why it started about loan. I mean, maybe do you have experience on it? And what was the rational behind it? You need to make a very vast with the large language model. You mentioned it. Most of our initial customers were loan officers were in the banking and lending industry. So we started off with a broad tool. What it would do is it would look at your emails and then it would respond to emails for you, sent client follow ups. And guess what? Loan officers send a lot of follow ups. So we got a stream of loan officers into the platform. And then we decided to niche and ask them, well, what more can we do for you, other than helping you write emails faster. So we did like hundreds of customer interviews, then we got use cases. And actually, I was still in school by then. So that gave me the free time to do all of these interviews without having an actual product market for the time. So once I graduated, I had a host of ideas to move forward with. OK. And now if we talk about a little bit on the technology side, how the it's processing a loan for applications, which was not possible before, if someone is using anti AI, so maybe give a context of if before, if someone is not using any if we start using it, what is the difference before and after. Right. I'll give context on what the technology does. I think that would be a good anecdotes to talk about how it helps the before and after. All right. So I think a high is the first truly non deterministic technology, you know, before when you give a computer an input, you know exactly the kind of input that was going to be, you know, displayed to you. So if you look at it from the context of loans, you have, let's say a W2 that you need to process. And let's say you put all of the numbers into an Excel spreadsheet. And then you did do like a sum. And then it gets you there, you know, total income and whatnot. Now, when you're looking at it from an AI perspective, there's a lot of risk of hallucination. So the AI model just making stuff up. There's also a lot of risk of privacy, security, sending models to open AI. So we looked at it from those two angles and we're like, well, we could train our own models and not only train our own models, but train our models for like specific things that caused the most bottlenecks for loan officers. So one of the things is when people submit a loan application, it is never 100% complete. You never like submit a loan application. You get all your documents. And then in two days, you're like, it's done. Here's your money, go buy a house. It's usually missing like one or two documents that you forgot the loan officers looking at it and find some anomalies and then calls you to request those documents. You might be at work that creates some bottlenecks. So what we did was we built an AI that would look at the entirety of the loan application, determine all of the documents that are missing. And then give the borrower a call requesting those documents and that doesn't have a map to it. Doesn't have like risks of like hallucinations. And because we really focused the model to be constrained. So that's one of the areas that we're targeting before you have to make a bunch of calls to borrowers. Now you don't have to make a bunch of calls, which saves you like a couple extra days in the loan origination process. Okay. So it's like reading the files, maybe with the with the rag or something. And then you're processing with large language models. And before reaching to the loan officer, you're trying to to finalize the application so that you can have a final application. Is this what I'm understanding. Yes, we do a lot of rag. So which we will augmentation. We also fine tune the models. And in addition to fine tuning the models, we make sure that the models are hosted on separate instances for each of our customers that way they don't like, you know, share data, we're hosted on separate endpoints. Okay. And how these are fitting within the larger ecosystem of banks where they have those financials and and the applications and the systems inside those. Yeah. So banks are really hard to adopt new technologies. And that's why I asked it. They have a 50 year old main phrase running cobalt and you're coming with the shiny new AI tool. And the only way it could be done. That's the way we're doing it is if you do a bunch of integrations with the tools that they already use. You don't replace any of their tools, right yet, you build on top of them. So in the loan origination industry. There are tools like encompass, which is sort of a loan origination system. They're tools like and sync code, which is sort of a POS for orders that upload their documents to the portal. And then tools like email and all of that. So with email, you could choose to integrate with the email client itself, or you just build your own SMTP server from the ground up and just ingest all very. So how it fits within the larger ecosystem is writing a bunch of custom code for customers. So when a customer comes in or like what are all of the tools that you use. And then we give an engineer and say go figure out how to integrate all of their tools into our software and then bring all of the data and then convert them into embeddings and then you know label those embeddings. We use a graph database. A regular embeddings would not really work for our process. And then that would really help match notes and vertices that when you search for something. It doesn't only give you the similarity, but it gives you like similarities to similarities to similarities to similarities like a bunch of similar churches. And yeah, it's a bunch of integrations and a bunch of custom code. And do you have now paid customers with you? Yes, we have a lot of paid customers. We have banks, one bank in Switzerland that's using it. We have independent mortgage banks in the US. They're using it. And next week we're actually doing our first Australian deployments. So that's going to be interesting. Okay. And you said it's a private installation for each customer. So is it just an endpoint or are the whole solution, including the fine to model your putting it separately for each customer. Yeah. So the front end that they see is a web app that's hosted on a specific domain. The AI models is an endpoint. So we train the models to deploy them to GCP vertex AI and then we put them into the separate end points and then we create those end points. And then we do the database integrations. And then I think some of the database stuff we try to make it like it's like completely separate instance, even like the authentication information like user names and passwords. It's only for the organization. And then after that, we come in and then figure out what are all of the tools that they use. And then we have to write some customers that doesn't really take a long time because we have really great engineers, but the front end that they see is is a web app. What I'm very excited about is the AI agent stuff because user interface for AI agents, the whole idea of AI agents is that they do something you tell them and they do something for you and actually with the results. But how do you know the process that they took, right? How do you know like the buttons that they clicked and things like that? It's a problem that we're figuring out. So some of our AI agents that your telephone calls, we just provide transcripts and then analysis of the transcripts so people can see it. But it's really hard for you to intervene while the AI agent is like doing stuff. So I wish there would be some sort of like CCTV camera just like looking at your AI doing stuff and then you can like, you know, jump in like in the middle of a call of like the agent doing a call and then you can, you know, tell them to to go a different way, but it's an exciting piece of technology and really excited for like what's going to come in like next, you know, four or five years. All right. Now, maybe if from our audience point of view, if someone is going through similar to your that they want to have the large length models, fine tune it based on their own data. And you have gone through this journey of maybe last one and two years and you have a now working product, which is being used by several customers. So maybe if you can give us some, if we maybe narrow down to top three tips or suggestions for them, how they can take it up in their industry, whatsoever. Right. Um, I would say. If I think about this question a little bit deeper, I will talk about my process and how we did it. So usually the market will pull the right product out of you. Um, so just follow the market. So when we started started with like an email assistance, like calling the chat GPT API and then even tragedy, you do not have an API at the time. They lost your API. My parents later on. So we reverse engineer the front end of tragedy to like get their access to a concern. Then usually you write a call. Once the API started, we call the church of T API and then a influx of, um, of open source models came in, then we started experimenting with those and we even built our own fine tuning library. But then other people started building fine tuning libraries. So the market, we're like, well, we can't compete with like an AI lab of 50 people building fine tune levels. Let's use their own fine tuning libraries. Right. It's like always following the iteration of the market, like keeping up to date with everything that's happening. One of the things that we're looking at right now is the recent old one models from at GPT that can do chain of reasoning. We haven't quite experimented with it yet, but it's moving so fast that you have to be agile enough to change it up to the market. That I would say is my biggest tip. I know you asked for three tips and I can go deeper. They're like other tips, but that's like one of my biggest Okay. So you kept yourself aligned to what what is coming and how, how you can take it up to iterate your process or maybe if you have developed something, take the new feature or are the way of doing it. And then you started infusing it in your journey. Yeah. Okay. All right. Now, and if someone is is just you know, a planning to do something on AI or maybe he's being forced as an as an organizational level or maybe due to market hype that you need to do something in AI in your organization. So is there any any framework or any views you can share with them that how they can start from being based on your experience. Yeah, I have a very specific framework on this and this is something that we actually went through. So a lot of people want to build general AI tools like they think I'm going to build a tool that would generate follow up emails for all sales people. I'm going to build a tool that would, you know, respond to, let's say, generate marketing copy for all people are going to build a tool that would, you know, be in like Google docs and and write documents for people. A problem with that is well, two problems. There's a limitation on the technology. It's not general enough yet. It may not be for the next 10 years. Second problem is a lot of the major use cases that are general enough. They have billion dollar companies, even trillion dollar companies working on it. In the case of if you want to build an email tool, Google Microsoft, they're integrating AI into Gmail and outlook. You want to build something with vision, um, mid journey is doing stuff with images and vision on the only way that you can compete is if you compete in an area where you have the most niche knowledge about and then there's no incentive for an incumbent to do. So that can be anything from like if you're into podcasting. I would say different types of podcasts like tech podcasts require different things from let's say marketing podcasts. So if you're a tech podcast or think of like the most narrow thing I can do in tech podcasting and then try to build a solution for that and build a solution for like one person. And then you start expanding from that build solution for like two people. It's technology is unpredictable that you have to be so niche. And then develop before you go out to market. That's that's a framework that I think think of something that you're deeply, deeply knowledgeable about and then try to build a very niche solution for before, you know, before expanding. Okay, now as this technology is emerging and it's a new right and so there are not so much practices are being followed for maybe from coding development to taking to find tuning and even even the rag and then how to take it up rag and how to evaluate it and self evaluate it and maybe manage it. So what were the your top challenges, which you faced as your team while developing this product and keeping in mind that you have a target to go and how quickly you can go to market. So maybe linking with your pressure that okay, I need to release it maybe six months and then the technology and everything's a new and and it's popping up so many things. Yeah, you know, the rise of the AI engineer people think like you have to know a lot of machine learning and stuff that is true. But 80% of machine learning is just software engineering. It's just like manipulating data, you know, doing evaluations and things like that. I would say, you need like one or two deep machine learning engineers. And then you need a lot of software engineers and AI engineers and one of the challenges that I faced was I studied software engineering. I took a couple of data science classes. I took a couple of machine learning classes, but I'm a software engineer hard. I did like distributed systems. And so try to bring on people on the team who have machine learning expertise and also try to learn machine learning when I need to know what's going on under the hood. And then few months into the company, we made a resolution that we would only hire AI engineers. So you don't have to be a machine learning engineer, but you have to deeply understand like, what is rag, what are embeddings, how does the neural network work under the hood, you don't have to. But you just have to understand like how it works. So hiring people who are knowledgeable about AI, but not necessarily machine learning. So I feel like there should be a separation of what an AI engineer is. What a machine learning engineer is and what a research machine learning researcher is. And this is three different vastly different things. And for company like us, we need a few machine learning engineers and a lot of AI engineers. Okay. So maybe this brought me to a question, which I usually ask and the rational behind is that I want to get some tips for the carrier who want a new carrier. So you must be interviewing a lot of engineers are in a different. So maybe some of the tips which you target to see in the candidates. Right. I have three things that I look for in a candidate. The first thing that I look for is, well, can they ship fast, have they ship something in the past. So one of my interview questions is, I asked them, you have one week sprints and you have two week sprints. Which one would you have the same scope, which one would you go for and why? There's only a couple of right answers to this. Some people would say, well, I like the two experience because it gives me like a lot of time to like develop and then test and then after that, I, you know, I can ship. That's okay, but we're company that's very early. We move really fast. Right. I like the two answers that I like is when someone says I like one week sprints because I can ship fast or someone says I like two week sprints because in the first week, I can ship fast and I can spend all the time in the second week testing the stuff that I've shipped in the first week. So that I like stage stuff in the first week and then test and second, that's like one of my tips that what I look for in a candidate. Second thing that I look for in a candidate is, have you published stuff on the internet just for yourself that no one uses. So I asked people this question, like, what are some of the projects that you have built that no one like it's only for you, but you still published it on the internet. You would be surprised people tell you, oh, hey, I've built a restaurant app and I'm like, okay, but like, is it? Do you have a restaurant? Like, do you have a personal story to tell. I had to reject a candidate who had built an amazing gym app, but it was not because they were into working out. It was just because they thought this is something cool to build and they, and then they built it. What you want in a startup is people who care about the best startups are people solving problems with themselves. So find people who build stuff out themselves and then put it on the internet. Then the last thing that I look for is what are you obsessed with outside of coding? This is the last question I asked. And I've got an amazing answer to that question. Someone told me I was obsessed with the ukulele and then they told me all of the ukulele records and then we went back to the people who are obsessed with something else. So these are the three things that are in a report and I recognize that as this go public anymore interviews that I do. I mean, you would say the strategy exactly and you're right. It happened with me in a previous role. I was interviewing maybe every day's two or three candidates and what happened some point of time, the typical my questions started, you know, I started getting very easily answers. Then I realized there is some fishy man and then I spoke to the candidate sent to the recruiters and then they said, you know, the people knows each other and then if you keep asking in the same typical questions three, four, tricky one, yeah, they talk to each other. And then I thought, okay, now let's let's bring the new, you know, flavor to the question. So that's true. But eventually the idea of asking this question is that if someone is preparing for them for for the next job or role, then he can get the mindset of an interview and who's hiring and hiring manager. That's the whole objective of this giving to that. Exactly. Okay. Yeah. Okay. So now if we summarize your journey of from a setback failures point of view, which one is the best failure you you are naming it. Yeah, that was a right or a best failure that that give you a head start to something else. Best failure. Link to the fast, it's linked to the house fast, you can ship it, you know, which which the someone is shipping very fast. Even like failure in terms of like my career journey or like failure and in the products. Anything you can phrase it, maybe it in a carrier, maybe it can be to learning some tool or maybe going for a features or maybe doing a market validation. Or maybe talking to a customer's and then you know, customers is okay. No, this feature is not really available. I don't want to use it. And you know, that's maybe I developed myself it's linking to that conversation. I think I'll do one for products and other one for carrier. So I think one for the product that I wouldn't really classify as a failure was when we built the email tool. Scaled it to over 20,000 users, but we did not win that market couldn't ship as fast as the other people and I did not have as much marketing intellect. So we got viral really early on, we even had Elon Musk retweet the launch post and then We got 10,000 signups in one week and then a couple more hundred after that and then really skilled to 20,000 until like other tools came up. And then that tool is sitting at over 10,000 users now, but I would not consider it as a failure, but I would consider it as like not dominating the market because when that thing happened, I think we're the second people that used chat GPT as a crew extension to write emails directly in your your the UI of your of your involves and So how that accelerated progress is it forced us to look for more specific niches that, you know, just happens to pay more enterprises and banks, they pay more happens to be, I would say, more impactful. With what we're doing instead of just like helping people where an email is faster now we help people close ones faster making people like happy and getting their dream homes and things like that. So that's sort of like a product failure that I think would would help. And one for a career. I think it will be It would be unfair to say that I have had like a lot of career fillers, which I definitely have. I mean, I've filled interviews. There was one time that I That I interviewed for Bloomberg and I'm African. So the person that I was interviewing with, excuse me, I have a little bit of a raspy throughout this morning. So the interviewer was it was an internship interview. And then when I got in there. I was asked a question, just like a traveling salesman problem, but it was like related to like flights and like picking the right decision for a flight. This then a bunch of flights to optimal routes, blah, blah, blah. And then because he was African, I was like, oh, hey, I heard his accent. I'm like, I can finesse and that probably like solve the problem, but like five with him. And then he ended the interview really early. I sucked really, really bad. Like it was one of the worst interviews I've ever been on. And then after that interview, I was like, well, how can I get better? So I made a YouTube channel and started making videos about algorithms and data structures and then made all of them public. So I would I found like hundreds of algorithms and they're still up on YouTube. I have a repository and GitHub that's public. Is filled myself solving these algorithms and data structures problems. So like for the last few years of undergrad, I solved a bunch of problems. And then once I was going to apply, after undergrad, because like that failure led me to build a lot of these like algorithms and do a structure questions. And I'm getting like three offers from Amazon, Snap and Twitch, you know, I think that is what career failure that, you know, propelled me to. That's that's that's that's some inspiring story, even from from putting it into and filming it and putting it in that. And frankly, when we started a conversation, you asked me about the audience, right, that age, I mean, I force myself to do this, this podcast and video or any blog writing and article. And I'm bringing from maybe like 2008 and mine and the only reason I do it, because whenever I saw, I see some new topic. And which I want to learn it, I force myself to go and do some half cook work and then I just go, okay, publish it. And that gives me the new, new information, right, that's the whole objective and that's keep forcing me because I'm not very regular. You know, because I don't have I have that some commercial or anything behind, right, the only reason is that if I want to learn something new, or maybe if I'm doing some day to day work. And okay, let's come and then write it down about that. That's a good story. That's that's the only, you know, I'm following it a long time, I'm following it. So I'm I'm happy with that. I mean, it puts you some time on pressure that, okay, you need to be extra conscious, how to write what to write. Even maybe you're not expert, but nowadays I don't bother. I just go right and publish. I have a question for you. So since you've been doing this since 2008, it looks like you've seen like a lot of. You know, trends like, you know, picks and local maximas and, you know, like things that people thought were going to be transformative technologies and turns out they were just a fad. I have two questions. My first question is, do you have a pattern? Have you developed a pattern in your brain for noticing technologies that would become better? And it's okay if you can't explain that pattern, but like have you discovered that pattern for AI in particular, these things are changing so fast. Where would you see AI companies in the next four or five years? Like what would you describe as an AI native company in the next four or five years? Okay. So let me answer it in a twofold, maybe as I mean, I was at I'm a technology guy and in my day to day work, I was a very hardcore implementing ERPs for for organizations and large enterprises for financial supply chain to HR, on all those type of, okay. And that started giving me a lot of a lot of insights that over the years you do with large organizations, what are the repetitive, you know, strategies coming from different, okay. They want to consolidate someone will come it and then they break it and then they goes back to again some holding companies, so that started coming pattern. Now in 2020, we were doing a lot of IOT block and and and primary block and IOT Web 3 came after that in 2019, we were doing a lot of demos, you know, everywhere block and every problem will be solved for the block chain, that's what we were doing as also. And now then at that point of time, I thought, okay, what they need to do is I need to do some structure program and in 20, I started a regular, the way you were saying it yet that you went into CS program. So I started, okay, I'm a full full time employee, so let me start with an MSI in AI in from University of Bath in UK. So I started in 2020 at a structural program around 20, half year, I completed last year and when I completed, I never thank you, I never read about generative AI in the whole program because it was not there, right. So all what we did was was machine learning algorithms and then social, critical and software engineering and all mathematics, Python and all that. And then when I came out from the university and then and that you rightly mentioned for an interview and I went into for an interview and then the guy asked me about the genetic way I am bad being and all that. And I said, dude, I just, you know, came out and now I need to learn, I don't know because it was not, it was not taught in even University. So those patterns, this is how you see it as a person now coming back to the question of an AI and how we're saying it, it's not like a blockchain. I mean, I spend so many to one, two years working with blockchain demos, preparing it, it is not an IoT and all that. So it's, it started impacting even our kids, even, even right. So 10 years, four years, six years or 10 years is asking about judgment and all that. And now it's being being added into each and every component of an applicant, either just smaller price or a large enterprise, right, so it started impacting it. So it's me, now the race has just began to wear to not to add an AI. I like I like that thinking. So, I mean, so I think we'll see as the world progress and then it comes back, but I believe the way your application is solving problem, my brother is doing his, he has his online this cake shop, which is a designer cakes for parties and birdies and all that. And he has his problem of okay, how I will write on recipe, how I will answer to the customers, okay, and how I will manage my detailed patients and I want one package, one, one AI coming from a taking on order and delivering it to that one. That's what you mentioned that solving the actual problem of someone that will lead to it. Now if that guy started thinking about it, okay, how I can bring something, you know, within my business, which is, which is ease or maybe increase my productivity. And that's where if you see from everywhere, we're starting seeing it productivity gains. Yeah. Thank you for asking questions. And okay, all right. So we have spent a lot of time. So maybe before we conclude our talk for today, if someone want to reach out to you, or maybe what is the best medium to share with them. Yeah. I'm pretty active on LinkedIn. So if you Google, well, Michael Vandy, search me up on LinkedIn. I'm the first link that comes up. I'm very active on there. If you want to have a call with me, you can go to our website at addy.so. And then if you book a call there, there is a hundred percent certainty that I will be in that call. So that's one way to just like get me on a call immediately, just book a call on the website. Thank you so much for joining. And how do you see addy.da moving forward? Oh, man, I think a lot of things are going to happen. We, the technology has the potential to skill not only from like mortgage loans, but from like auto loans, car loans, a lot of things, anything that requires like manual document review. I think that is something that AI would would handle. And we're doing a whole lot of work in the document review process, like converting like unstructured data to structured data that manipulating that. So super excited forwards to come. Thanks for listening to the O Tech Talks podcast and be part of tech talks at otechtalks.tv. It's a turf to share ideas, insights and innovations.