INTRODUCTION

Castle is a Silicon Valley startup with its roots in Sweden, Malmö. The company provides SaaS products focused on account safety and fraud prevention. By constantly analyzing vast amounts of user behavior patterns, Castle lets customers act on risk scores based on their own preferences. The scores are computed based on statistical patterns, determining the probability of a single or a sequence of events to be malicious.

CHALLENGE

We were tasked with designing and implementing Castle’s state-of-the-art AI-pipelines, allowing the company to rapidly iterate on new models from idea to customer facing production environments. This included setting up distributed computing tools, machine learning model lifecycle and building dashboards to manage models, datasets and clusters.

GOAL

Enable the company to create, develop, test, deploy, monitor and maintain AI models for live anomaly detection in time series data with sub-50 ms response times.

SOLUTION

We implemented and deployed multiple services and cloud offerings. In essence, we wanted to be able to service model predictions in under 50 ms. This required clever use of caches, precomputed features and scalable infrastructure.

Training the models relied heavily on querying and running through large amounts of data. We deployed and managed distributed computing clusters (Spark, Dask) on AWS and ran multiple jobs to train models for fraud prevention. This included modelling using outlier and anomaly detection, spectral clustering, smoothing methods and Bayesian statistics.

We built an engine to run tasks in the cloud. Tasks were jobs, either dataset preparation or model training runs. In addition, an API was built to manage and handle models, datasets and clusters. This was presented in a dashboard.