Gyrus has several models for different applications, use cases and verticals. Gyrus engages with customers in a timely and phased manner to deploy machine learning models. The nominal timelines for each of the phases is as below.
The engagement starts with a collaborative effort with customer to define the use-case. At the end of the definition phase, the problem statement, acceptance criteria is defined, the data sources, available formats are identified, and the deployment model is agreed upon.
Gyrus goes off and prepares the data and develops the model. At this stage, the data quality is measured and all the data enhancement techniques are implemented. If the data quality is very low, this stage may take longer. Gyrus models are taken and transfer learned to the specific problem identified.
In the next stage, the model is deployed and monitored for performance and usage. This is a feedback loop stage where minor tweaks are expected. A/B testing can be employed to validate the new models’ efficiency.
The final stage is roll out with the confidence of the results from the deployment.
In the development process Customers come in with the problem statement and Customers’ data set. Gyrus comes in with our ML/AI models, hardware platforms and standard datasets.
In the next stage, the customer data is cleaned, enriched and made ready for data science. The models are developed with transfer learning by starting from one of the prebuilt models.
The corresponding explainability models, and bias checks are performed on the model in addition to the differential privacy checks and enhancements.
Mckinsey says 25% of all the AI models need continuous improvement. It can be with new data, new features or requirements. Model maintenance is key for reproducibility, but a challenge, as version control of code will not suffice. The datasets, algorithm implementation, parameters used need to be version controlled to be able to reproduce results.