Skip to content

Experts in fintech-specific AI

The financial tech sector is our bread and butter.

Experts in end-to-end development of fintech specific AI

The financial tech sector is our bread and butter. We have developed solutions in multiple contexts over the years and are well versed in the specific challenges of the domain.

The founding team has played an integral role in building machine learning capabilities for some of the most successful fintech scaleups in Europe. We all worked at Klarna where we built their first credit models, machine learning-based fraud detection systems, and the models responsible for deciding optimal payment options for customers using the checkout service. Our legacy includes building credit models for Qliro, a local competitor to Klarna, as well as various risk management components for Collector bank. Since the inception of Modulai, we have been a central partner in various Fintech projects in both Europe and the US. We have advised and developed conceptual models for Trustly, one of the largest fintech companies in Europe, devoted to online bank payments. Together with Corpia Group, we have taken on the difficult task of estimating company risk for lenders focusing on the SME segment. Our solutions for company default risk prediction are among the most innovative we know of. And consist of multiple modules predicting bankruptcies as well as sophisticated time series models for predicting the future macroeconomic climate for the affected sector. We even advise Riksbanken (Swedish central bank) on our findings around the use of advanced natural language processing models for describing the current economic climate.

Challenges specific to fintech

Input data and regulations
Input data in financial sector applications needs to be properly understood by the engineers involved. They sometimes come in the form of various ratios, or quantities, referring to certain posts in financial statements. This is especially true when the system will estimate company risk. Furthermore, indicators of risk are received by external third parties and denote characteristics with regard to local regulatory frameworks. These frameworks need to be interpreted, understood and sometimes they pose hard limits on how they are used for the process to stay compliant.

Models sometimes are required to be explainable, meaning that the decisions they make are human-interpretable through the contribution weights of the input variables. Historically, this has limited modeling approaches to strictly linear ones, where the model output is directly related to weighted sums of the input. Lately, however, interpretability has become a highly active field within artificial intelligence and quite successful attempts have been made at making high-capacity machine learning models and neural networks interpretable in a similar fashion as in the linear case. We at Modulai have extensive experience on the topic and we believe that the tradeoff between model interpretability and predictive accuracy is less of an issue today. We’re certain that with the correct tools we can open the door to more creative model development within the fintech sector.

Realtime systems
Many AI use cases in fintech need the models to run in real-time environments to produce predictions at the point of transaction. Precomputed model outputs are often infeasible, if not impossible to use since model inputs and the decision policies require data unknown beforehand. Examples of factors that affect this are, for example, the nature of the user and the transaction itself. Dependencies on external data sources such as credit bureau and KYC/id-services pose another level of complexity due to external API calls that have to be completed prior to any model inference being run. All in all, one might end up with a multistep system with dependencies on both internal data sources as well as external ones, each consuming execution time, and a total response time of around a second. On top of all this, the system needs to be scalable and able to handle Black-Friday-level of throughput, be easily maintained, and be cost-effective.

We have developed machine learning-based decision systems for fintech for years, some of them with very high demand on robustness, speed, and cost-efficiency. We have faced tough design decisions, tight deadlines, and challenging performance requirements. At the end of the day, the systems we’ve built have made millions of credit and fraud decisions in real-time. If you have been buying something online in the Nordics, your transaction is likely to have been processed through one of our systems.