BlueCloud and LendingTree Live with Snowflake

Bill Tennant: Everybody, thanks for coming. I am fortunate to be joined by two people who know this content inside and out and will hopefully add much value to the discussion today. I'm Bill Tennant, Chief Revenue Officer of BlueCloud. We also have Scott Taubman with us. He's the Chief Product and Technology Officer of LendingTree. We will talk a lot about the work we've done with LendingTree and Snowflake's professional services today. I am also lucky to have up here on stage with us Chase Romano, Senior Architect in AI and ML, Solution Innovation with Snowflake, who genuinely helped make this project tremendously successful across the board.

So, with that, we're going to take a snapshot of LendingTree, its business, some of the complexities, the initial challenges it faced during its migration, and how it arrived at producing data science workloads.

I will explain why they chose Snowflake and the process they went through to ensure they had the right future-proof solution. I will also mention the advantage that BlueCloud provided and how we worked better together with Snowflake Professional Services and great people like Chase.

From there, we'll discuss our current state and innovative plans for the future. So, I will hand it off to the team and let them take it away.

Scott: LendingTree, just as a snapshot, is a two-sided marketplace. The easiest way to think about it is that one side of the marketplace is the product for the other side of the marketplace. So, consumers come in looking for a financial product we're trying to provide, and our lender partners are looking for quality consumers. We're trying to give each side of the marketplace the best version of the other.

I was joking with my colleagues when I made this slide, thinking it looked so easy and that my job should be easy. However, no two lenders are the same, and no two consumers are the same, which creates some challenges. And then, we had to do this in real time, which also created an interesting set of challenges. LendingTree started as a mortgage company, but over the years, personal loans, business loans, credit cards, and a whole bunch of business units have all followed a very similar model.

The company started at Internet 1.0. Everything was on-prem, and there were lots of proprietary software solutions. Even when the company went to the cloud, it didn't leverage many managed services. It was treated as a virtual data center. It repurposed many technologies and used them in unnatural ways, which also complicated our lives. The financial crisis brought the company to its knees.

COVID did the same thing, but then interest rates dropped, which was a boom for the company. As interest rates went back up, the ReFi market evaporated overnight. So, the company was trying to navigate all of this, which created technical debt as time went on, without much opportunity to focus on it.

The company went on an acquisition spree to diversify itself and insulate against the economy. So, we created our own technical debt, then bought other people's technical debt and put it all together. But the businesses we bought all work slightly differently. For example, just the way they model revenue—some of them focus on revenue per lead, and some of them gain revenue straight on a click.

They're all nuanced and used different terminology and lexicon across the data and architecture, creating a mess generally. And what did that wind up looking like? We didn't have a single source of truth, and for those unfamiliar with the term, it's really the way for the company to look at the same data in the same place in the same way.

What we had instead were a bunch of different data silos from different business units - a lot of tribal knowledge about how things work, how things are organized. Trying to combine that into a unified view for the company was very taxing on the system. So, we had a lot of performance issues.

Some reports took over five hours to run if they didn't encounter issues, and we had to run them overnight. We couldn't share much more than simple data with our partners. We wanted to be a good partner for the other side of the marketplace, but we were giving them the basics of leads and revenue. The most fun part was the circular references, where we had a database looking to another database for information, which in turn was calling back to the original database to get the information. So, lots of fun technical debt to work through.

Why does this matter?

Over time, consumers had higher demands, lenders wanted better performance, and competition crept in. Going back to the triangle, now, as a more attractive circle, it's getting better at this. Having this flywheel model where the more we know about a consumer, the better that consumer looks to the lender creates a positive feedback loop and positive outcomes.

So, just getting a credit score isn't good enough anymore. You want to know what the propensity is to transact. Customers may need something else even if they came for a personal loan. They may need a credit card instead. For the lenders, we wanted to know which consumers are closing deals and which ones are going to get a rate lock on the mortgage, which, for us, is a perfect sign that they're going to close deals so that we can send more consumers that look like that to that lender. At the end of the day, the desired outcome is just giving each other, each side of the marketplace, the best product possible.

So why did we choose Snowflake?

To take a step back for a second, changing out your analytics platform is like a heart transplant. You're trying to keep the patient alive the whole time you do this, and we are doing it in a very economically challenging environment.

So, when our interest rates went up, the ReFi market disappeared. Then, I asked for seven figures to make a Snowflake investment, and I could not explain why to the CFO. They wanted to know when I was getting their money back.

But what we were looking for was future-proof architecture. You're making a 10-year-time horizon bet when you're doing something like this. Anything short of that, you won't get a return on it. So, you want to make sure that it's here for the long haul. What's the strength of the partner ecosystem? Who else is on here? It's an open-source community - you want to go where the developers are and ensure the code base is thriving. The same thing goes for the vendor that you pick. This included increasing the set of services that Chase will also talk about.

Finally, it had to be cloud agnostic. We acquired a huge insurance company. They were in Azure, and we're in AWS. It made sense to stop and perform that cloud consolidation, so we needed something that worked across AWS and Azure. And Snowflake fit the bill. The only challenge is that we didn't have any in-house talent who knew Snowflake. Snowflake is a fundamentally different architecture requiring a very specialized skill set so that we couldn't navigate it on our own from the start. And we needed to bring in some outside help. So along came BlueCloud.

Bill: Again, we are lucky enough to have the opportunity to work with the great teams at LendingTree and Snowflake. How does BlueCloud operate? We call it a Borderless Delivery Manifesto. By leveraging talent markets worldwide across dozens of different areas, geographies, countries, and capabilities, BlueCloud has given us the ability to bring in premium talent at very cost-effective rates and provide opportunities worldwide.

We genuinely believe that we're making a difference, not only within our company but also to the end consumers that we're working with. Everything we're doing is focused on developing future-proof design, making sure that we're targeting specific use cases with end outcomes that are valuable to the businesses that we're working with.

At our core, we are an elite Snowflake partner. We have competencies across six industries and support hundreds of individuals around the globe. We're small enough to be nimble and meaningfully support our customers.

But also, we have the talent, we have the diversity, and we have the capabilities that you would generally find with some of the larger GSIs out on the market. Overall, being able to provide data at its heart, and then all the things to the right and left of Snowflake, such as the ability to execute software engineering, UI design and development, AI and machine learning, overall migrations, and the ability to help with source systems, like Salesforce across the board allow us to support companies through and provide value.

Scott: We were also looking for pricing flexibility. LendingTree was going in without a lot of money, really no budget. BlueCloud has this nice service where you can pay for highly specialized talents and have them on-site sitting next to your teams you can pay for. If you're okay with somebody who works potentially off hours doing more mundane tasks, you can do that offshore at affordable prices. This was helpful for us when we had to prove the value initially and then start increasing our investment there.

We wanted a good partner - not too small, not too big. We wanted a partner who could be flexible with the company changes. Our budget and priorities changed, and we needed to put this project on hold. BlueCloud navigated this with us; we're great partners along the way. And then we also wanted Snowflake to say, "We know BlueCloud, and we bless them." That's a big one. You want to go with somebody who Snowflake recommends.

So, what goals did we want to achieve? As a CTO of LendingTree, I'm looking for three things. I want to reduce our technical footprint. I want to modernize our tech stack and future-proof it. I also want to give developers and analysts a platform on which they want to work. The folks you want to hire want to work with these technologies. It's better for retention, hiring, and a competitive market. They want to be working on today's technologies.

Our transactional databases on the left still capture form submissions in real time. But in parallel, it's all streaming in Snowflake. So, if anyone in the company asks, "Where's my data?" The answer is, "It's in Snowflake, and it's in one place."

We were able to offset the cost of this by skipping the transactional databases altogether in some cases. We tried to do real-time analytics on production databases, which was a bad idea. We were able to put those away, move them into Snowflake, and do that in near real time, which was a big win. And we're able to integrate with partners now. CoreLogic is a good example of how you normally sign a deal like that, call your APIs, and pull data in. But they're in the Snowflake marketplace, so we can integrate them quickly.

All right, so what are some of the early results we saw?

As I mentioned with the databases, we didn't fully offset the cost of Snowflake, but we made a really big dent in it. By retiring large, old, legacy data warehouses, we saw a 40 percent cost reduction in our cloud spend. The 5-hour unreliable report I mentioned is now done in 10 minutes.

Also, we now make decisions on that report. It used to be a once-a-day decision. For example, we used to look at the Google ad keyword spend at the end of the day and decide to change that tomorrow. Now, we can do that hourly. By being able to make those decisions in near real-time, we can adjust to them and save a lot of money over the course of the year.

This increased visibility. Now, we have account-level visibility for who's using the platform and who is spending the most on the platform. We have a high score for somebody who wrote a 15,000-dollar query in Snowflake. It was a very poorly optimized query. We now have a wall of shame for people who are being too inefficient with their Snowflake spending. So that's a good way for us to keep track of our Snowflake spending in a fun way. So we want to make sure that the spending is taking place and there's a corresponding ROI behind it. Chase, anything to add on there?

Chase: We see this a lot with customers who need visibility into their spending. We can track every single query and know who's running it. We can set the parameters to shut queries off if necessary. And then there are the data sharing capabilities. I was just joking with some of my coworkers earlier—I would do a huge presentation on AI and ML, and then someone would just ask me: "Wait, you were able to just get a copy of that person's data without having to do file transfer or any of that?!" That's a really big piece because, as a data scientist, I believe it's not just about the model; it's about feature engineering and what features I can get as a data scientist. And, at LendingTree, the more features, the better.

We were able to get tons of features for this algorithm that we built with BlueCloud. I work at Snowflake, and I can tell you that we have customers talking about it. But the ease of getting all those features in, being able to do that feature engineering, and building a model that will drive ROI —that's what I strive to do as a data scientist.

It's not just a nice PowerPoint. Scott and the CFO of LendingTree, they want to see dollars. They want to see ROI. So being able to work with BlueCloud on this project and being able to work with other data scientists and build this model especially with you guys was pretty awesome.

Scott: Alright, we'll get straight into that. So, the previous slide shows some of the savings and efficiencies that will not cut it for the CFO. They want to see ROI. And so, selfishly, my job was to go out and prove that we made the right decision and that we could leverage advanced analytics in a way the company couldn't previously.

We had hired data scientists in the past, and they quit because they couldn't do anything. The challenge we've been trying to solve for years is this idea called capacity smoothing. I mentioned earlier that we have lenders, large and small. Some lenders can only take ten leads a day. In the past, the way things would work is that they could get those ten leads between 8 and 8:30 in the morning, and then they're done for the day. And, for a small mom-and-pop shop, they can't make this many phone calls and follow up. And so they're not getting the value out of their network.

We wanted to get a lead per hour, considering time zones and different segmentation of the consumers, as well as who works best for this small or large lender.

Some people want a lot more time on the phone, and others don't want to be bothered by it. We try to learn as much as we can about the customer and distribute leads evenly over the course of the day. For instance, now, at 4 p.m., our partners haven't hit their capacity for the day, and they can still take leads as they come in.

So that tees up the problem. And we brought in Chase to help solve this one.

Chase: Yeah, and it was an interesting problem I've never faced before with another customer. It's not just a financial problem. It's a customer matching a lender to a lendee. We need to figure out how not to overwhelm our smaller lenders because they need the employees to take in those loans. But we also don't want to leave money on the table for them and LendingTree. This is a mutually beneficial model. So, we also need to figure out what lenders we can hold off on in the morning because they're bigger lenders.

They will get leads later in the day, and we won't overwhelm them in the morning. We may slowly smooth it out. To do this, we need to know everything about the customers, and we need to know data about the lenders. So that's a lot of data. Three years ago, this was not possible in Snowflake. When I first started at Snowflake, we couldn't even run Python. Now, you can train a model in Snowflake. You can do all the feature engineering. The data science team at LendingTree was already leveraging VS Code and SageMaker, so they'd already had experience. And with Snowflake, we do have our notebooks, but they could just work in VS Code, connect to Snowflake, and write Python.

One of the biggest things in data science customers were doing, and what LendingTree was doing before, is leveraging a third-party tool. AWS is a huge partner of ours. A lot of people are using SageMaker. But to leverage SageMaker with Snowflake, where must that data go? It has to leave Snowflake, go to SageMaker, build the model, and send it back. Many egress fees are incorporated, with lots of data movement and security. You also have to provide data access. Now, with the ability to run Python and build AI and ML models in Snowflake, we can leave the data in Snowflake and train large models on it. We had hundreds of extra features added to it.

It's a very wide table, and we're able to use Snowflake's compute power to scale up for the training, scale back down for the smaller task, and leverage that model to distribute the leads more evenly throughout the day. So, keeping everything in one place will also save a lot of money.

So, if you're moving data out of Snowflake into another data science tool because you think you need it, you can do that all natively in Snowflake now.

Scott: And this was a six-month plus effort?

Chase: Yeah. There was a lot of feature engineering and trial and error. We had to do a lot of A-B tests in situations like this because, let's say, we put this model in production, and it doesn't smooth out. That's money lost, right? Every time we're running a new iteration of the model, you have to test it to make sure it's actually going to drive revenue and not just smooth it out and make the lenders happy; we need to make the lenders happy while also driving revenue. We had to find that equilibrium between keeping the lenders happy and offering them the right financial incentives so they'd follow through with those loan requests.

Scott: You scared me initially because the first model was dramatically less successful than the model we had in production. To the broader point, it was important to solve this problem and create an easily repeatable pipeline. So, depending on which model you go with, you want to make that a very easy experiment to run. We went through dozens of models before finding the right one while pulling in different data sets. And now not only have we solved this problem, but it's a seven-figure improvement to our business, which is fantastic. We also now have a pipeline where we can solve the next problem much faster. We invested with BlueCloud and Snowflake to do this the right way and create a repeatable pattern.

So if the next experiments pan out, it's not a huge investment lost. It all comes down to a lot of experimentation.

Chase: One of the greatest things about working with BlueCloud is their independence and ability to solve problems without guiding them. My team and I work with many customers worldwide to get workloads in production, specifically AI and ML. When we started working with LendingTree and BlueCloud, it was nice because I didn't have to retrain BlueCloud, which I have to do with many SI partners. After all, things are growing so much in Snowflake. We have a lot of new features coming out in Snowflake, and it's really tough for the SI partners to keep up. I need to learn them, and then they will have to learn them.

What was neat about this project was that I was able to hand off work to BlueCloud, and they were able to run with it. I didn't have to overshadow them or babysit them. I could give them materials and show them a new feature quickly, and then they could go and show the feature to the rest of their team.

I would come back the next week, and it'd be done. That allowed me to work with a second team at LendingTree and get a different model in production while BlueCloud was able to handle this massive model that was being done. We were able to work together and get multiple models out running because BlueCloud was able to take this and run with it. The SI partners work with internal teams at Snowflake, where together, we make sure that you're using the best of the new features while also getting guidance from your SI partner.

Bill: It was quite an experience. One of the first things you and I discussed was how we could save you on costs. And then Scott said, "I've done as much as I can by proving my value by saving money; I need to drive dollars to this company." This is one of the first initiatives where we took a holistic approach to driving revenue. And I think it's created a foundation for growth.

Scott: When we brought in BlueCloud, there was a lot of resistance within the company. We had contracting companies that we had worked with. We didn't want to lose those people. Now, I have folks asking for more BlueCloud blood, which is a good validation point for me and the relationship we've built. So that's fantastic.

I would also like to share some key learnings and key takeaways from this migration. Please learn from my mistakes.

Bring in the experts early. It will help you finish things sooner, but it's also almost side-by-side training as you build up the internal talent within your company. Show a path to value, even if you can't execute it immediately. You need to show there is a light at the end of the tunnel. Bring the tools to the data. This is a big shift with Snowflake. As Chase said, taking your data out, putting it into another tool, executing it, bringing it back, that's going away.

On the other hand, putting the tools on top of your data because data is getting so large and so painful to move around is going to be the new model for everybody.

On the downside, please don't migrate your technical data. If you have a mess of a data platform like we had, it will take you twice as long to do it the right way, but it'll be well worth the effort.

Everybody keeps asking me when we will be done with the Snowflake migration. And I keep saying, that's the wrong question. The question is, "When are you getting value from it?" There's a long tail of silly little databases that will take us another year to get off, and it doesn't matter to the business.

By the end of this year, the business will be 100 percent running off Snowflake, and then we've got a mess to clean up, which should be done under the radar.

Chase: What I want to talk about here is predictive analytics. Are there any data scientists in the room? The ability to build these models now is just so much faster. When it comes to actually doing predictive analytics, I know there's this huge push for GenAI, and people forget ML. I just want to point out that there is no GenAI in this project.

Think about that ML side. Don't just focus on GenAI. With Snowflake, you can move your classic ML workloads to Snowflake now. Being able to run predictive analytics at a scale will save you tons of money.

So, in this use case, we're running inference as these customers come to this website. We will have that prediction in real time, match them, and decide whether we want to send them to this lender or that lender. All of that can be orchestrated. Think about how you can bring your predictive analytics to Snowflake, and I guarantee you that if you are running it on an outside platform, it will probably be faster. You will be able to do it the right way if the data is there.

Your data scientists can engineer features and iterate models. One of the biggest things about being a data scientist is that you must be able to discover and accept the truth. You've got to be able to call your baby ugly sometimes.

Throw it away. Stop trying to make it better. You've got to know when you have a bad model. We had that. We could say, "Look, this iteration, this model is not working. Let's throw it away. Let's start over." You've got to be able to do that.

As Scott said, it took a few iterations, but once we found the one that worked, you could run it every day and generate ROI. And being able to do that quickly and efficiently, where your data sits, is just beneficial.

Scott: We can talk a little bit about what's coming ahead. Now that we've got the foundations in place, there are a whole bunch of business opportunities that companies want to execute. And we finally can do that can. It's a good place to be. The partner ecosystem keeps getting stronger, the number of services is increasing, and the number of investments Snowflake is making in this area is staggering.

--

 

Ready to unlock the full potential of your data? Connect with BlueCloud and Snowflake today and start transforming your business with cutting-edge solutions.