How startup Capital Float is using Machine Learning to automate loan processing

Capital Float, an online platform that provides working capital finance to SMEs, has leveraged technology to build a differentiated model that is able to deliver credit to the smallest of businesses in a scalable and efficient manner. Rohan Angrish, CTO, Capital Float, tells EC, how his firm is leveraging innovative techniques such as leveraging cellphone penetration to offer customized financial services, APIs to achieve higher levels of scale and efficiency, and machine learning to take internal efficiencies to a new level

Some edited excerpts:

How do you use technology to differentiate your firm?
Today, the data and technology today has reached a point where we can actually bring down both in cost to the originator as well as pricing for purchase-taking loans. We see ourselves as setting up a new trajectory or setting a new paradigm within which to operate where other players will come in and participate. We call this coopetition. We are looking at being not being a lender but being a lending platform. We create this differentiator by capturing as many data points that are possible about the customer.

Lets take the example of SMEs to explain this concept. If you break down the segment of SMEs across two different axis, there are online SMEs and offline SMEs. Within this online segment, there are SMEs about which you have a reliable digital footprint of some shape and size. There is some reliable digital data and then there are SMEs out of which there is a more reliable digital footprint. Now the e-commerce guys fall squarely in their online businesses and there’s a reliable digital footprint that we can use to underwrite that particular business which is why it’s a sort of a low hanging fruit where Capital Float entered. And we penetrated that market to the extent that we became the leading player in the market. We set the trend and we set the business model .

Then you have the other quadrant which is offline businesses about whom we have no digital footprint. That’s hard to crack, that is the kirana who is sitting in a Tier 3, Tier 4 town or maybe in a village. So the Holy Grail is to effectively get to those people but there is this sort of transitionary middle quadrant which is really interesting ich is an offline shop or an offline SME but has a digital footprint and this is where Capital Float deals.

Now if you take a look at a kirana store, you may have come to that store to do top-up for the cell phone or recharge a prepaid SIM card. So there is some amount of that business where we can get reliable third party digital data about this business. Now, in this case, the kirana store doesn’t have an intention of going online to sell its products. However, it still has a reliable digital footprint. We identified this opportunity, and we launched our kirana app. A kirana store can download the Capital Float app and apply for a loan using purely Aadhaar. At the same time I’m connected with this remittance player at the backend and get transactional data about the kirana store on that platform. I use this transactional data to create some sort of partial picture of what this person’s business looks like, at the same time also got permission to scrape this person’s SMSes. From SMSes, we can reconstruct how they top up their phone.

So using all these things, we were able to create a credit profile for a kirana store that otherwise had actually no credit profile. No one has ever looked at that kirana store to create a credit profile, but by giving them our app we created a credit profile. By putting a bunch of pieces together, we can give kirana stores a loan offer on the spot with our automated decision engineering loan app that is running in the background . So within minutes of opening our app, SMEs can have money in their account. We are able to give a kirana store money in minutes completely digitally, paper free in a regulatory, compliant and safe way. So whether it’s a kirana store, whether it’s the handicraft market, whether you have a self-help group or even individuals, we can have everyone mapped.

How do you leverage APIs?
Today standards are built in a very one-to-one way where the APIs are built and ready to use just one partner. If I bring in another partner, then I have got to build another API. To overcome this issue, we are trying to come up with standards for data exchange while being regulatory compliant. So it will need things like ‘permissions’, it will need things like consent in the form of digital consent, but at the end of the day, the idea is to come up with platforms where people just come in and plug in either as producers of data or consumers of data.

We will become front runners in the cashless economy with credit as a hook. Companies are signing Capital Float and saying that I want to enter the market where I want to give credit as a hook, and I want to partner with you.

How do you leverage machine learning?
When we took our first call in the market, it was necessarily intuitive because we had no data to back up. We were taking the models that we knew worked somewhere else, and hoped that the models will work in this system and we started. We tweaked the traditional model a bit and started drawing data from the system. Data started flowing back to us. We started pushing cash in one direction but what that means is that data starts flowing in the opposite direction and we started learning from that system. As I started learning from that system, I then was able to — at a certain point, tweak my model, and push more money, more efficiency into the system and then that led to a more virtuous cycle. The more loans we pushed, the more data flowed back into the system, and better were the loans, as we kept on learning. All of these cycles of iteration, today within Capital Float, are all based on machine learning. There is no human being sitting and trying to figure it out.

What are the efficiencies as a result of machine learning techniques being applied?
There are certain loans today, certain loan profiles within Capital Float today that are done end to end by a machine. No human being looks at the loan profile. So we only start tracking the loan performance when the loan starts coming back in, in which case we have an early warning system that helps us track a potentially, a loan that might be going bad so that we can get in touch with the customer and ensure that the loan does not go bad. But the idea that there are certain loans today which are within our risk appetite and within our level of comfort and we have learnt enough to know that there is no human being needed over there. It’s kind of like a cockpit of a Boeing 747, where the plane can fly itself, even though it doesn’t need a human being, there is still a pilot sitting over there. In case something goes wrong, the pilot gets you.

Earlier we did not have this capability, but today a significant part of our small ticket sizes is completely human free. The small ticket segment is completely automated end to end in terms of risk profiling and underwriting. But that being said, there is no part of Capital Float that has zero automated risk profile and underwriting. In fact, in the worst case, the machine does about 80% of the work and the human being has to do the remaining 20. It even gives a suggestion to the human being on what must be done. For example, if the ticket sizes and loans are outside of the ability or comfort zone to let the machine take the call itself. That’s when the human being comes in and does the remaining thing to make sure that everything was done correctly. So in the worst case, we have 80-20 split and in the best case machines are doing 100%.  We will be able to go from small ticket sizes to slightly bigger ticket sizes. Over a period of time, we will be see automated underwriting grow.