“Ever since the beginning of the pandemic, the complexity of frauds has gone up many folds,” said Parikshit Chitalkar, co-founder of digital lending platform StashFin.
Founded in 2016 by Tushar Aggarwal, Parikshit and Shruti Aggarwal, StashFin targets young urban grey collar employees. The platform has acquired a million customers so far.
The digital lending industry is projected to reach $20.31 billion market cap by 2027, growing at a CAGR of 16.7 percent between 2020 and 2027.
StashFin works like a neo-banking platform providing credit line cards powered by VISA and MasterCard. “Think of it as an overdraft facility banks provide,” Parikshit said.
The platform uses technology to facilitate customer onboarding and providing seamless experiences. A potential customer has to follow a five-step process to register on StashFin. First step is to share the picture of their Pan card and Aadhaar card. Then, the ML model extracts details to fill the registration form. The user has to enter an OTP, post which the underwriting model triggers in the background, and decide:
- Whether to approve or reject a certain user for a credit line
- Amount of credit a user is authorised for
“The system uses close to 3,000 variables to analyse a customer before announcing the decision. These variables can range from bureau score or credit score to analysing how desperate a user is for credit, the time at which they are applying for credit, their banking history, etc,” Parikshit said.
StashFin uses basic models and standard APIs to examine the validity of official documents. The platform asks the user to click a selfie with the documents provided and uses a TensorFlow-based image recognition model to cross verify. “Doing this is not easy since the Pan card image would have been taken somewhere at a fairly young age, the selfie would be recent, and the Aadhaar card would have a picture clicked somewhere in between,” Parikshit said.
StashFin’s classification model raises a red flag if the image is not consistent. Typically, fraudsters try to confuse the ML model in two ways:
- When the user clicks a selfie with another picture of themselves at the back, thus, confusing the model by having two images in the same frame
- Wearing a t-shirt or jacket with a face printed on them.
In such cases, the model either provides a wrong result or returns a null value. When returned with a null value, StashFin does a manual check.
“We also have to ensure that we do not get false positives which might result in a good customer getting mistakenly detected,” he added.
For this, the digital lending platform does backtest on customers with good credit scores and are only a couple of EMIs away from completing their payments. “We backtest on that population since the fraud rate or ID theft is next to zero. If it throws a single positive, we try to work to minimise it,” Parikshit said.
Once the credit is allocated, the ML model figures out any anomalies in the pattern of usage. It uses up to 35 different micromodels to analyse and predict usage patterns. Some of these models are also self-reinforcing. Each time a user makes a payment, the model looks at where they stand in their credit cycle; if they have paid on time or not. The ML model makes decisions based on a user’s payment history, like reducing the interest rate for people consistently paying on time.
Finally, StashFin uses ML models to help users make informed financial choices. For instance, the platform guides users to nearest ATMs, and stores where Stashfin users can redeem their reward points.
StashFin has built all its tech in-house. It has an engineering and data scientists’ team constantly working on improving the product.
With a million customers on-boarded, StashFin plans to increase the number by up to 7x in the next 18 to 24 months. “We want to be sure that our platform is ready to handle that kind of scale,” Parikshit said.
Soon processes in the fintech industry will be digitised and KYC technology is in for an overhaul. Many paper processes have to die to make engineering efforts to automate, monitor and alert processes, he added.