Interview: How A Medellín-Area Tech Firm is Helping its Global Clients Compete With AI & Data Science

In 1900, industry ran on coal. Today, business runs on data. Competitive advantage comes from having the data, but more importantly, knowing how to extract value from it. With the rise of AI, everything everywhere is getting disrupted, all at once. Those companies, governments and individuals who have the resources to deploy artificial intelligence are destined to out-compete the laggards.

One company that is finding rapid traction supporting international businesses, especially those in North America, is BI Consulting Lab, located in the residential Medellín suburb of Guarne, Colombia. The company’s offices are a beehive of activity, with data scientists, programmers, and client executives buzzing about. Company founder Brian Castaño Rendón made time for Loren Moss, the executive editor of Finance Colombia and Cognitive Business News to come by and learn more about the firm’s rapid progress.

Finance Colombia: So I’m here with Brian Castaño Rendón, a friend, a colleague, a former co-worker. I’ve known you for quite a few years now, probably seven or eight years. Cool to reunite with you here at your company headquarters in Guarne, which is a bedroom community outside of Medellín. And you’re doing some really interesting things. Tell me about your company. What’s your company called?

Brian Rendón: My company is called BI Consulting Lab. First of all, thank you for having me and for spending time with us.

Finance Colombia: Of course. So, what is BI Consulting Lab? I know it’s software-focused. I know it’s data-focused. But tell me in your own words what BI Consulting Lab does.

Brian Rendón: OK, so basically, what we are is a company that specializes in data and AI. And we use that technology to help businesses make better decisions. We do decision support. And we try to allow the executives and the managers of every company to have the tools to come up with the decisions and support those decisions. Hopefully, by implementing the right decisions, the productivity of the company will progress exponentially over time.

Finance Colombia: I think one of the things that happens when I think about the companies that I’ve worked in, small companies, large companies… A lot of times, the problem is not collecting the data or not having enough data, but not being able to extract the value out of their data. I think that what happens is we might have, I don’t know, let’s say a grocer who has two types of inventory. They have the actual products that they sell, but then they have their shelf space. And it’s easy to say, “How many bottles of ketchup did I sell today?” or “How many cans of tuna fish did I sell?” or something like that.

But then there’s another level to that. And that’s to say, “How do I optimize my inventory when it comes to real estate?” And I don’t mean real estate like buildings or shopping centers, but I mean the shelf space that they have, and how they might not know how it is going to affect sales if I move this product from one level on the shelf to another level. Or “How do I optimize the shelf space?”

One of the things that annoys me here in Colombia, I’m more of a Pepsi guy than a Coke guy, but I go into the store and Coca-Cola has like eight racks of just Coca-Cola. And there’s like two bottles -not two racks- there’s like two bottles of Pepsi. Now, they’re not doing that by accident. They must know something. I do know that Coke is more popular in Colombia than Pepsi, but maybe there’s more to it than that. Maybe the way that the stores optimize their space, they’re going to not just put things on the shelf because they want to, but they have to figure out what the optimum size is.

What are other use cases for where a company might want to call on you all and enlist your help, not just as software people, but as real consultants who can go in and analyze and look at what they’re doing?

Brian Rendón: Yeah, of course. So, to your point there, when you were talking about specifically the retail industry, we do have experience working with large retail stores, specifically in the US. We are working right now with Kroger Acosta, which is a supplier from Kroger. We help them analyze their space distribution, their shelf distribution, and what they call planograms.

Let’s take three steps back. And, what you said first, companies don’t have issues collecting data, not that much of an issue, all of them have spreadsheets. All of them have Excel. Most of them already have SQL databases and stuff like that. What they struggle with is extracting its value. And usually here in the data industry, we say that you usually need three things to be able to extract the value from the data, to convert the data into information, and the information into decisions.

You need three elements to converge: you need the business knowledge, you need the technical side that allows you to do the computations, and then you need the statistics. So you need the ability to concatenate these three elements, like in a Venn diagram. And at the center of it, you kind of have the formula for you to be able to extract information and convert it into decisions.

So, on the shelf space, just to take an example, in this case for Kroger, what we did was to create something called a space analysis tool. What we did for them was to help them get all of those hundreds of millions of records that they have on the different commodities and the different stores and the different departments and sub-commodities and divisions, and stuff like that, organizing it into a cohesive model. When we say modeling in data analytics, we mean the relationship between dimensions and transactions for a company, and how we can take a specific transaction and dimensionalize the transaction into when it happened, how it happened, and why it happened. So we create a cohesive model, take those hundreds of millions of transactions, and we try to put the values into perspective.

And if we alter this or that combination, we could, with probable cost, assume that we could get a better performance, or maybe we could have a worse performance if we do that. So, having the data and organizing it helps you to do those what-if analysis scenarios that allow you to make the right decisions or try to make the more probable right decisions, because this is a game of probability.

Finance Colombia: Crystal ball or something.

Brian Rendón: Crystal ball, exactly. Thank you. Yeah, so what we deal with here is probability. So if you do this, data-supported, it could probably go up; if you do this or that, the data tells us that it probably won’t go as well. And then we do a hypothesis, we test it on our models, and then we give you a possibility and a scenario that you could utilize in your decision-making. That’s an example of how businesses, specifically from the distribution company, the distribution and the retail company, can come to us and we can try to help them improve their decision making, for example, in their distribution or their shelf space in a specific market chain. But that applies to any industry.

That’s the cool thing about data, because data is omnipresent in all industries. There could be problems whether they capture it, all of it, or not, but they are constantly being generated at every time in every industry, in every division of any company. So what we try to do with that is to take it, analyze it, capitalize it, and utilize it in a way that the company can make decisions based on that. By improving their decision-making, they can improve their productivity.

“Companies don’t have issues collecting data…What they struggle with is extracting its value.”

Finance Colombia: I think one of the areas, I look at forecasting, and I think even to stay in the grocery sector, I remember I’ve been down here in Colombia for a long time now, but I remember right before I left and moved down here, a really interesting thing was happening in the US at Christmas season. And that was that, traditionally you have Christmas, and then right after Christmas, you have these big giant clearance sales to get rid of all the Christmas stuff and all the stuff left over from Christmas.

Well, in the later years, those were going away. And the reason… I mean, there’s always some sale, but it wasn’t because they had excess inventory, but because the forecasting got so scientific that they were able to put in all of these variables and economic indicators and all this stuff like that. And they were able to predict very, very accurately how much of whatever product they would sell. So by the time Christmas was over, they didn’t run out before Christmas, but there wasn’t an excess inventory after Christmas.

And I think about how these types of tools and disciplines that you all specialize in can be applied. Do you guys have any cases where you’re able to use your data expertise for predictive analysis, let’s say, or to be able to do forecasting for clients?

Brian Rendón: Yes, we do that all the time. Where we focus is on trying to understand the historical trends. And for that, machine learning is a pretty powerful tool to do that. So when a company has a lot of records and a lot of, let’s say, let’s think in a tabular way. So you have a table with your transactions, and those transactions have a lot of characteristic attributes. So you have the type of store, the type of commodity, three or four subdivisions, they have dates, they have comments, even, like plain text, string elements. And we can use all of that and put it into a deep learning algorithm, for example, an algorithm.

And we can try to understand what attributes are the most impactful about the transactions to discover patterns or clusters, or different elements that allow us to understand the data. And then what we do based on those algorithms is to take the most important elements for us to predict a consistent behavior from the historical data.

So we first allocate the important elements for us to do a calculation, and then we set up the calculation. And what we do then is do an 80% to 20% analysis. So you take 80% of the data to train a model, and 20% to test it. So when that 20% of the data, when the model gives you a performance, or I would say around 98%, 97% of accuracy, then you take that information that you got left, that the algorithm has never seen, you fit it to it. You already know the answers for each of those values. But you want to test how accurately they predict the right price for a specific day, for a specific week, maybe, or month. And you want to see how the prediction compares to the actuals.

And if you see that there is not much difference, you have a statistical base to say, “Okay, based on this statistical way that has proven 80% to 90% accuracy, let’s plan our inventories by that, and let’s keep testing it.” So we do a six-month, three-month, and a year forecast, and we even adjust the model to allow us to adjust the forecast as time goes on. Because as time goes on, the model has more data to be trained on, and has actual factual data that has happened with the respective attributes, and can infer the most important elements for the whole bunch of time you have on your data source. Or you can even segment it, segment it, and then try to find asynchronous spikes or stuff like that.

Finance Colombia: How far away, and I’m sure that there are so many other business use cases for what you guys do, but you guys also are very well-versed in AI, in building AI into your data solutions. How far away are we from, let’s suppose one of your clients says, “Okay, we have AI and we have all of this data, and can I just ask the software, hey, how much shelf space should I allocate to Seven Up?” In plain text, just like an AI, like a generative question, or is it still a lot more complex than that? Even though, yes, I know we can get the data, but how far away are we from being able to just say, “Should the Tide laundry detergent be on the third shelf or the second shelf?”

Brian Rendón: Yeah, so I would say on that one, we’re not that far, we’re rather close, but we need to figure out a couple of things. So let me go into that. So when you or your users use ChatGPT or Claude or DeepSeek, what they do is to enter a context in the context window, the chat window that you have, you start typing a problem or a comment or whatever, you give some instructions, you add some context, and based on the training that the model has given, he will answer you, right? That’s basically how it works.

The issue with doing that with corporate data is that, first, the corporate data is not that well-structured in most cases, so that’s the first thing. And then most of the models limit your ability to input information into the model. So the way it works is that you buy, for OpenAI, which is the most famous one of them all, you buy a subscription to them, right? You connect via an API, and you insert information that you want into the model so it can use it as context and then answer you. So each of those lines, each one of those syllables that you input is are token, and for those tokens, you are charged money.

So it becomes very inefficient when you want to process millions and millions of records, which will turn into billions of tokens, which you will have to pay for to get a single answer. So that is not an efficient solution right away. So, for example, what we do to try to improve that and make that process more efficient is we do the connection to the model, the LLM, the Large Language Model, even Anthropic or Claude or OpenAI, in this case, ChatGPT. We can connect to all of those models. We can connect to the open source, which Facebook is the powerhouse for that; we can connect to Llama. You can connect to all of those models or any of those models.

Brian Rendón: For you to make it efficient, you want to have like a small step in between. So what you want to do is you query your information directly via an API, so that the API provides you with the actual value. And then you use the LLM, ChatGPT, to generate a text-based answer. So it tells you in a normal English-speaking way the answer that you want to hear. But the processing, you push the processing not into the model, but into the SQL engine, or the database engine, or the data warehouse that you use.

And there are different ways for you to do that. There are also ways you can input multiple inputs into the model, so the model processes them and generates an answer.

But usually the most effective way for easy questions, for example, if  you are in finance, you say, “Okay, how many subscribers did we get last week?” You can go and check your database and do a query, and get an answer. But if you want to talk to an assistant, we can do that. We can connect your WhatsApp, for example, to an LLM that will connect to your database. And you ask that, that will go to do a query into your database and get an answer. And that answer will give you back your answer either by text or by speech, whatever you prefer. If you send a speech message, a voice message, it will answer you with another voice message. We can improve the voice using Flask, another model that is out there. So we can do that. And in that way, the LLM, the AI model, is not doing the processing of the information.

Finance Colombia: It’s just packaging it.

Brian Rendón: Exactly. It’s just doing the delivery nicely.

Finance Colombia: Like in networking, we used to call that the presentation layer.

Brian Rendón: Yeah, exactly. And that’s important because the way our LLMs work at the moment is that they get confused. When you give them a lot of information, like you overextend in your context window, the LLM model will start to hallucinate. What I mean by hallucinate is that it will start to give you a random answer that is not correct. And that’s because the model itself doesn’t know what answers are correct or not. They answer in a probabilistic way. And that’s precisely why you don’t want it to process millions of records of information, at least not at this stage. Because it can get confused. And you don’t want to take an answer from it, maybe a hallucinated one that is not accurate, and then you allocate some of your budget based on a not very accurate answer.

Finance Colombia: Exactly.

Brian Rendón: So, at the moment, at least, you try to push the processing with accurate models that we can easily do at the moment using APIs. And you query the information from the model. First, you need to have a very good data architecture, which that’s some other thing that we do. We help businesses to understand how to capture the information correctly, how to extract it, normalize it, and place it rationally into a data warehouse, Datamore, and different databases.

So that their transactional information is arrayed with their dimensional information, so that they can be combined in a way that is efficient and answers questions for you. So when you have a very robust and very good data architecture, it facilitates a lot the transition to implement AI models. Because when you have accurate information, and rather than accurate information, very well normalized and organized information in your company, that information can be easily queried.

And when you query that information, which is easily, again, queried, and you push the processing into a very powerful engine like SQL, for example, you get specific answers, so the right amount of answers that you want. And then you put it into the AI. And then the AI can just give you the packaging that you set, tell you the answer nicely, and it can perform a couple of other layers of analysis as well.

So you can ask it to give me a projection of sales for my next six months. So it pushes into the SQL engine a query for your actual revenue or whatever metric you are measuring for your last 18 months, and uses that as an input for it to calculate your following months. And it can give you a trend. It can give you a chart. It can give you a lot of things. And it is based on actual information that, with the right amount that it can process with a high degree of certainty that is accurate.

Again, we are still working on the probabilistic element. The different models right now, what I believe they’re doing based on the news that we hear, is to make more efficient what we have, to reduce the number of hallucinations, to make it have short-term reasoning and long-term reasoning, which is basically what we do. We come up with an easy answer, but then we reflect on that answer and maybe confirm the answer. Or in other words, change the answer.

Finance Colombia: Refine it.

Brian Rendón: Refine it, exactly. So if you see one for OpenAI, the ChatGPT model, that one usually gives you, when you ask it something, it starts reasoning. And you can visualize the questions that it is asking itself, and how it conceptualizes the question that you asked, and the possible answers, and try to refine it. It was way lower, like eight months ago, but it gives you more accurate answers nowadays.

Finance Colombia: What are the applications that companies can use that you’ve seen in your business, and with the kinds of companies that you’ve all been able to help? Where companies have engaged you to help them improve their customer experience. How can the work that you do translate into a better customer experience for companies? Do you have any examples of that?

Brian Rendón: Yeah, we have a couple of them. Usually, in the context of virtual engagement, for example, they have multiple challenges to be approached. And they are struggling to be able to answer, one person to answer in different channels, keep track of the different requirements, and answer them in ways that are more optimal with the whole context at every time.

So it’s hard for them to keep track of all their chats in WhatsApp, Instagram, and maybe others. So what we do for them is we create automations that connect up their context database. So we create a context database for them. And then we connect that context database to ChatGPT, for example, to OpenAI or other AI models. And those AI models get, we’d say we create a custom ChatGPT, trained for that specific business. And that way, whenever there is a person who comes, that triggers a reaction in our pipeline, that pushes the question or the comment from the social network as an input, and then sends it to the LLM that has already been trained with the context of the company. So they have the context on how to answer that question, or comments, based on whatever the company has at the moment. So it is very important, the connection, the connectivity with the databases of the company, at every turn. Even more so when you are trying to do a service on time, and live services and stuff like that. So we have a couple of restaurants with that. And we are discussing a solution for a stockbroker here in Colombia. They… I met him at the event that you held at the Marquee.

Finance Colombia: Oh, great. Last year. So, Finance Colombia events lead to business.

Brian Rendón: Exactly. Yes, it does. So he told me that time, “Look, I’m the CEO of the company. We are a very large stockbroker company. And whenever we have, let’s say, the dividends coming from Ecopetrol, it is a madhouse in my company because we have 13,000 customers calling at the same time trying to confirm that we are depositing their dividends, on how they went or how much it came, stuff like that.” And he told me, “I have 15 people at customer service and they get overwhelmed immediately with that type of information.”

So I thought at the moment a couple of ways to reduce that pressure from customers trying to get access to their data specifically. I asked him, “Okay, we could do dashboards for them, dashboarding and reporting, custom reporting for them that they will automatically get once the transaction comes with the usual context that they ask for, so that they already have it, and that way they get this incentivized to call because they already have the information and that could reduce the number.”

Another thing is to try to automate part of your customer service using these LLMs. There was a more recent solution that we proposed to him, because he said people want to talk to someone, to hear, and since we are advancing pretty fast in the speech element of the AI, the sounding is less robotic, it can be trained.

Finance Colombia: It’s getting pretty spooky. Yeah, it’s getting pretty impressive.

Brian Rendón: Yeah, it’s getting pretty impressive. Today we can connect to those models and train it into a Colombian, Argentinian, Cuban, Mexican, whatever type of accent that you want, it can be pretty realistic and can help you send automatic messages or answer calls if you need to and give the answer, the simple answers to people based on a couple of parameters. So you need to parametrize all of that.

Finance Colombia: You’ve been very generous with your time. Last question for you. So we’re here in Guarne. Guarne is kind of a bedroom community, a suburb. We’re probably about 20 or 30 minutes out from the city limits of Medellín. You’re Colombian. I realize that, but why does Colombia work for a business of this type? If we think about your clients that are in the US and other parts of the world, a lot of times, people think of IT, they think of a company being in India, or something like that. Why Colombia and why the Medellín metropolitan area for a tech company of this type?

Brian Rendón: Okay, so I’ve been working with you actually for a lot of years now in the BPO business, and understanding the relationship, the US relationships in terms of outsourcing and in general, and specifically for technology. I know that there are three major hubs for outsourcing technology services. So we have Latin America, we have India, and we have Eastern Europe, right? So, when you have Latin America, I would say two elements may differentiate us from the other two. First is the time zone. We have a very good time zone.

“I would say Medellín stands high as a digital nomad center. We have a lot of people coming from different places. The last time I checked a nomad list, I think it’s called Nomad List. We were ranked, I think, first or second in Latin America.” – Brian Rendón

Finance Colombia: We’re in the US Central Time right now, Texas, Chicago.

Brian Rendón: We have the same time zones as the US, so that’s a first element. And then due to our proximity to them, we are more culturally close to them than an Eastern European from Ukraine or Belarus, or Macedonia.

Finance Colombia: Very true.

Brian Rendón: Or also an Indian. Our accent is usually more familiar to them. Our way to speak to things that are relevant to both sides is also closer. So there are ties, cultural ties, that facilitate the communication. So I would say that that’s something that maybe can be impactful for a decision-making process for a US company to choose a Colombian or Medellin company over a Belarusian company. That could be some major points there. But the other element that I would say is that we have… specifically Medellín, which is particularly high-skilled in technology than other areas in Latin America in general.

I would say more than Lima, more than Monterrey or Guadalajara, more than maybe Montevideo. I would say Medellín stands high as a digital nomad center. We have a lot of people coming from different places. The last time I checked a nomad list, I think it’s called Nomad List. We were ranked, I think, first or second in Latin America. And in the world, here in Colombia, we were like third or fourth after Bali, Barcelona, maybe another one. So we have a lot of talent coming over here. We have a very friendly environment for development. We have cultural ties, like I described before, the ability to work in the same time zone, the cultural ties, and the high-skilled people.

So that, I think, makes a pretty good combination that differentiates us from the Indian and the Belarusian. Also, we are a bit more expensive than our other Eastern European and Indian counterparts. But I think the quality part, which is hard to quantify sometimes, but I would say in most cases, the quality part, which is the more humane part, like the easy to speak in a way that is more understandable for people, can be a very good point.

We are very, very present in most of the operations. And within the Latin American people, Colombia and specifically Medellín stand pretty, pretty high in the rankings of people. Wherever I go, there are half-Latin American teams. I usually see people from Colombia, and specifically from Medellín.

Finance Colombia: It’s kind of like the Silicon Valley of Latin America. It is, you know, access is easy. We’re probably half an hour from the airport right here. And then once you’re on a flight, it’s two hours to the US. And yeah, I think that, you know, you also have that Paisa entrepreneurial culture that you find specifically in this part of the country, maybe a little bit in the Coffee Axis down in Coffee Country: Manizales, Pereira, which is part of this region. It is, you know.

So Brian, what is, if people want to get in touch with you, if people want to get in touch with your company, what are your if you could remind us of the company name and, of course, your website, but any other channels of communication that would be convenient for you and potential collaborators or clients?

Brian Rendón: Of course. So, first of all, our webpage, again, is www.biconsultinglab.com. My corporate email is [email protected]. You can also find me on my LinkedIn profile with the same address, Brian Esteban Castano Rendon. When you go to our web page, you have a Contact Us page, and that will automatically help you to set up, to direct you to a landing page where you can set up a meeting, a 15 meeting with one of our sales representatives. So we can do a first assessment, a quick assessment on whether we could be a good match for whatever you need.

This one may be important for people to know. So, the way we classify a company that can be a potential customer for us, we don’t receive anyone, because our solutions could help anyone, but not all companies are prepared, or at the stage where our services can be of good value for them. So, we are pretty sincere from the get-go, and we tell them, “Okay, look, we want companies that already have some sort of infrastructure of data or have enough transactions that they think is necessary for them to start setting up a data infrastructure.” We can help them there, where you want to make sense of your data, we can help you there, and when you want to start automating and maybe do predictive analysis and try to incorporate AI into operation, we can help you there.

Either you know what you need, we say, okay, maybe you specifically need this. We both agree, we give you a quote, and then we start our payment process, and then we do an onboarding, and we start working together.

Or if you want to kind of implement the concept of what we sell, like getting the data, if you have to improve your decision-making, but you don’y know how to do that practically, what we do is we offer you a free diagnosis, we take around three hours to make with your top executives and department, head of departments. And it has to happen within a week. So, we take the three, maybe four hours, we sit down, we understand your business, we understand your applications, the integration within the application, how you extract information from them, what are the pain points that you have, each of the departments in general, and also each of the departments, what are your strategic goals?

And we try to combine all of that information to understand your current state and where you could be. Once we understand those two, we lay out a plan that says, okay, maybe from, coming from where you are to where you could be, we think that you could do this and this and this or that steps, and we can help you do those, for this amount of money. And we will send you a formal offer. We did a diagnosis that, again, is free. So, you get your diagnosis at the very least, you get a good sense, a better sense of where you are, and how you could eventually implement AI or data in your day-to-day operation. And if, maybe, if we see that we could be a good fit and you like our proposition, we start working together, we discuss the quotation, you pay us, and we do an onboarding session with you. And hopefully we can bring you value, we can start a long-term relationship, helping you do the transition from a business model that is not AI or data-centric into one that very, very much values their data assets and can monetize them along the way.

Finance Colombia: Impressive, you know, and it’s been impressive to see your success grow as a person, you know, I’ve known you for almost 10 years, but, you know, we’re here in your offices, and it’s full of people. I walked in today, and it’s just like a beehive, everybody’s on computers and having different calls and things like that. So people must realize these aren’t just a couple of guys with a laptop; this is an impressive operation here.

So I would encourage, of course, any kind of company that is looking to better utilize their data to support their corporate decisions, or not even just companies, but even organizations, government agencies, things like that, that you guys obviously, you have the advantages of being based here in suburban Medellín, Colombia, but are already serving clients in North America and throughout the world. So, I’m excited for you, I’m excited to watch you guys grow, and of course, we’ll continue to keep in touch, and I can’t wait to hear about future successes, so congratulations.

Brian Rendón: Thank you so much.