We bring expertise in consulting, technology and analytics. As such we take a holistic approach to our client engagements, understanding your business, priorities and objectives as well as your data. This allows us to identify real opportunities and set them into the context of a broader digital strategy.
While neural network based models and NLP feature strongly in our solutions, descriptive statistics form an essential part of our upfront data analysis and statistical analysis of model performance. We also use time series analysis and modelling to identify patterns in seasonal data and sometimes use this to impute data, increasing datasets for ML applications. Bayes and other techniques, including XG Boost are used for early model prototypes.
Model development and production applications have very different characteristics. Production run-time environments must be stable, reliable, have predictable performance and high availability. Data analysis, model experimentation and development require flexibility to give the analyst the opportunity to quickly turn around prototypes. This is not conducive to a controlled production environment. Also, model training and evaluation, especially with deep learning models is computationally intensive. These intensive operations require GPU and other high-performance computing tools and can accumulate significant cost in cloud-based platforms.
We overcome these competing requirements by implementing two environments: a cloud based production environment with full DevOps control and a separate open, flexible and high-performance on-premise computing environment for model development, evaluation and training.