It started with designing a reliable storage layer capable of processing and storing a huge volume of transactions consistently. The custom ledger and supporting data stores enabled keeping track of millions of inflight web and e-commerce interactions that served as valuable data points for the underwriting. A dedicated event store provided the ability to store data points around these web and e-commerce interactions in time series fashion. All of this data without PII provided adequate insights into the online buying patterns of customers. With the foundational systems in place it was possible to access the information needed to make decisions. Most peoples go through their standard cycles of buying groceries, food, clothes, booking cabs and and paying for utilities in usually predictable manner. Having a regular merchant to buy groceries from, favourite restaurants, or paying the bills at the start of the month. This usually works as a good source of leading data points to identify the eligibility of an individual towards a purchase. The location, average daily, weekly and monthly spends aggregated over time provided for a decent insight into an individual’s capacity to spend. The hard bit was organising these data points to generate a reliable model. Simple data pipelines which crunched large volumes of data and funnelled them into models, maintainable by small teams proved effective. The cold start problem of having no information about potential customers was addressed with a Bayesian approach by dedicating a portion of the funds to customer acquisition which was used to take small bets on customers over a short windows to transact on trusted merchants to build guiding data.