Machine Learning Company

The company struggled with slow development velocity, stemming from architectural anti-patterns caused by a simplistic modelling approach, which resulted in tightly coupled code and coordination challenges between teams. Innovation was stifled by a legacy machine learning pipeline, a lack of MLOps control architecture, and a training data approach that required significant manual configuration and on-site customer support. High costs related to bare metal installations and upgrades further exacerbated the difficulties, leading to resource-heavy processes and customer dissatisfaction.

We helped a medium sized machine learning company scale their technology base and resolve development issues originating from the start up phase of their business.

To resolve slow development velocity, we advised the company to adapt its current system by refactoring the existing codebase, employing patterns from Domain Driven Design to rationalise data modelling . This approach avoided the high costs of switching frameworks and reduced the complexity of managing data across subsystems. To improve their machine learning pipeline and MLOps, we proposed using a general purpose workflow engine to enable more flexibility in customer configurations and experimentation, while transitioning to cloud-based infrastructure to handle installation and upgrade challenges more efficiently.

The company embarked on disentangling their model architecture, adopted Docker and Kubernetes deployments, and saw significant benefits in using a containerized workflow engine for machine learning experimentation, though they continued to face challenges with bare metal deployments.