6.6.24: MIT STEX Demo day - Catalan AI

ISHAAN GROVER: Hi, everyone. I'm Ishaan. I got my master's in machine learning from the MIT Media Lab, and I'm now finishing my PhD in the same lab. I'm also the co-founder of catalan.ai, which is a B2B SaaS dynamic pricing as a service company. And I know that can be a little bit of a mouthful, so I'm going to spend the next four minutes explaining to you exactly what we do.

But before I go on, I have a quick question. I just bought this banana today morning for about $0.15. Why does it cost $0.15? Why not 12 or why not 16? And why does anything cost what it does? Why does this shirt cost what it does? Why does my laptop cost what it does?

We were very curious and we asked around, from the smallest to the largest businesses in the world, about how they price. And almost everyone said something along the following lines. So we'll do cost plus pricing, which is I make my product for $100, I want 15% profit, so $115. Or they'll do competitor tracking, which is someone will say, I'm the market leader and I set the prices in the market or that I'm always going to be the cheapest in the market. So $0.01 below the nearest competitor. But these are all high level rules, not really driven by data.

But the reality is that the demand of any given product depends on a lot of factors. So it could be time of day, day of week. Is it a demand increasing holiday or a demand decreasing holiday? Weather, seasonality. And perhaps the most important, which is where are my competitors pricing today?

Now, the high level rules and strategies that people use, they don't really take any of this into account, which causes a lot of inefficiencies in the market, leaving a lot of money on the table. So imagine the inefficiency in pricing one product on one day and multiply that with all the products that you're selling. And that's now a significant portion of money that's left on the table.

So at Catalan, we decided to solve this problem once and for all. So we use AI and machine learning to find the right price of any given product at any given time. And what we do is we take vast amounts of historical data and we train our models to get an output that looks something like this. So on the x-axis, you have price. And on the y-axis, you have revenue. And what this graph tells you is how much you would have made at different price points.

So this is actually from a real merchant. And you can see that the cost was about $20. Competitor one was pricing at like 21.9. Competitor two was at 22.5, and the merchant currently was at about 22.8. And what our model said was you should price below competitor two but above competitor one, and that will yield an additional revenue lift of about $80,000. So this merchant did that and they actually saw these results.

But we didn't just want to create models in our workspace and then go on. We actually wanted to test them out in the real world. So in the last one and a half years, we've done many, many, many, many different pilots. But I'm going to focus on one specific use case where we did this pilot with one of the largest gas retailers, oil and gas retailers in Mexico.

And the way we did it was we took 12 stations. We made six control stations and six treatment stations. The control stations were naturally priced by the company, as always, and the six treatment stations, Catalan priced. So we actually changed the price on the billboard that you see outside a gas station. And after seven months, we generated a profit lift on the treatment stations for about 32% and a revenue lift of 17%.

But we didn't just stop there. We then went into groceries, where we generated a 15% revenue lift. And then we went into e-commerce, where we tried a lot of different industries. So apparel, fast fashion jewelry, vitamin A and vitamin C, the kind of products you would find at a CVS. Flowers, notebooks. And aggregated, we generated about 27% profit lift in e-commerce.

So if any of these problems resonate with you or if there are other AI and machine learning problems, we solve a lot of different things also. Or if you're in one of these industries or some other industry that's not listed here but has a similar problem, please reach out to us. Again, I'm Ishaan, and it was a pleasure meeting you.

SPEAKER: Thank you, Ishaan. Thanks so much. So what are some other problems that you might be trying to solve with your technology besides what you just explained?

ISHAAN GROVER: Right. So one of the biggest problems that we're seeing and opportunities where there's a lot of inefficiency is in deals. So a lot of these large companies have hundreds of different coupons and deals that they'll give out, but they're not personalized to any one person.

So effectively their usage is very low, which means if I wanted, let's say, a water bottle, if I don't see a deal for the water bottle, I'm probably not going to be able to use it and I might go somewhere else. So they have the infrastructure, but no machine learning or models in place to actually be able to give the right deal or discount to the right person. So we are right now beginning to find that.

We're also doing something in buy now, pay later, which is you want to set the right price and credit price to be able to maximize whatever objective the company has in buy now, pay later. We're also doing replenishment. So there's a lot of different use cases of the same technology.

SPEAKER: Great. Thanks. And how do you intake data to create these models for pricing predictions, and how much data do you want to have to be able to use effectively?

ISHAAN GROVER: Yeah, so that's a great question. So usually two years worth of data is very good, which is two years of transaction data. So all we need is what price did you set and how many things did you sell and what your costs were. So with that, we then augment with a lot of different things that we pull from the internet and then we create our models. So two years worth of data. But if you have less, we have worked with companies with six months of data and still been able to provide results.