ML Model Improvement & Management using MLOps
90% Improvement In ML Model Run Time Using MLOps
The customer is a leading multinational CPG company that produces health, hygiene, and nutrition products. They used the pricing and promotion model as an essential part of their sales strategy to attract potential customers. However, the model ran in only one country, and migrating it to multiple geographies was a huge challenge.
The company also missed business deadlines because of the long model run cycle. Moreover, the model was failing for multiple SKUs across retailers because of varying data. The customer also faced technical challenges such as poor prediction quality of the model, and huge time to model training that took up to 8 days to run. Some of the other challenges were:
- High cost of infrastructure for every model run
- Zero test coverage
- Non-scalable codes
Sigmoid developed a solution using MLOps that reduced the time to train the model and automated the model training. In this first phase, Sigmoid developed a test suite and automated model deployment and testing to improve the prediction outcomes.
In the next phase, we created benchmarking suites to check for model improvements and efficiently handle errors to further improve the performance.
In the last phase, model features were improved by identifying the ones that were failing and working with the machine learning team to create a rule-based approach to fix them. After the final rounds of testing, models were migrated to multiple countries in a consistent and unified way.
The new solution brought scalability and standardization and enabled the migration of the model across geographies to create a unified approach.
8 Days to 14 hours
reduction in model run time
reduction in cost per run
improvement in MAPE