Arkose Labs is a leading security fraud and abuse prevention organization, based in Brisbane, Australia, and San Francisco. They solve multimillion-dollar fraud problems for the world’s most targeted businesses with zero friction to users.
Rapid-growth saw their enforcement challenges being served to millions of users per day. To scale their analytical capabilities, and to improve the user experience, they have introduced machine learning models to help identify legitimate human-traffic, as well as significantly increasing challenge difficulty for bots and invalid users.
Arkose Labs partnered with ThoughtWorks, to enhance their Data Engineering and Machine Learning capabilities, using agile practices to build a scalable and production-ready Machine Learning framework, which increased the productivity of their Data Scientists. We implemented tools for managing research, feature engineering, dataset versioning, batch processing, and artifact management.
The framework needed to run with minimal operational overhead, so we opted for a serverless solution which provided sufficient automation for Data Scientists to run the platform themselves. We built a series of tools that gave us the flexibility to run tasks for rapid experimentation, as well as for robust recurring batch tasks.
Together, we integrated Machine Learning capabilities with existing monitoring tooling, so that Data Scientists could make decisions about when to retrain their models and when to investigate abnormal traffic patterns.
To complete the platform, ThoughtWorks delivered a series of web-based visualization tools, built with React, to help test and tune Machine Learning models. These tools used model interpretability approaches to provide Developers, QA’s, and Data Scientists a window into the performance of their live models.
Incorporating Agile software delivery into the data science and engineering space, as well as Continuous Integration and Deployment practices, allowed the team to respond quickly to the needs of the business, and reduce the risk of a problematic model deployment. The platform successfully delivered four new models to production, with an inference time low enough that Arkose Labs could handle all forecasted traffic with room to grow.