At the Women in AI conference, The Rising 2020, Tanvi Keswani and Sai Keerthisree V.N., Data Scientist and Data Engineer at Ericsson respectively, talked about one of the most challenging parts of AI often called ‘The Last Mile Challenge’.
The speakers discussed that the increase in AI has led the organisations to now focus on making data-driven decisions although they struggle to integrate the AI solution with existing business processes. Currently, the external world focuses on how to create a model and often forgets on preparing how to deploy and manage it.
The agenda of the talk included the overview of the machine learning cycle, approach to AI-powered demand forecasting, development and deployment of the pipeline, its challenges, among others.
The talk was started with understanding the business, formulating the problem statement, how to approach the problem and build the model and then discussed the deployment, how to package the AI solution, building data pipeline, containerization, versioning, and continuous integration and deployment.
They highlighted the end-to-end journey of an AI solution, that is, how one can make AI implementable and the challenges faced during deployment which could have been avoided in the development stage itself. They explained in more detail by walking the audience through a business use case from Ericsson Global Ltd.
Tanvi started the session by discussing the overview of the machine learning cycle, that includes various steps such as data collection, analysis, model, evaluation, validation, model deployment, monitoring, reporting and other such.
Tanvi stated, “Model deployment is not just a stand-alone model or a Jupyter notebook or a pickle file. Starting from collecting the data to industrialising it, there are many do’s and don’ts that one needs to be aware of and the amount of effort that goes into this is generally underrated.”
She then discussed the background of the demand forecasting business model at Ericsson. At Ericsson, the demand is volatile and complex to predict at the granular level. The manual forecasting model is a low accuracy lengthy process as well as time-consuming and struggles to capture market dynamics.
These issues have been mitigated by improving various forecasting steps of the current process. Data-driven decision making helps for faster and responsive supply chain management.
At Ericsson, improving the demand forecasting accuracy helped the company in yearly cost reductions through three key levers, which are-
- Inventory
- Logistics
- Scrap
According to Tanvi, the AI-powered demand forecasting at the company holds the following crucial steps-
- Scope
- Forecast for radio, baseband products
- Forecast at the lowest granular level
- Short-range forecast
- Model and Result Validation
- A feature-based approach to generate the forecast
- A back-tested model with significant improvement over manual forecast
- Desired Solution
- Automated forecasting engine
- Forecast available 2 months ahead in time which gives sufficient time for inventory management
- Value Impact
- The data-driven demand forecast could bring MSEK annual cost savings through reductions of inventory, freight and scrap costs.
After the discussion of how machine learning forecast yields improved accuracy and lowered bias compared to the traditional forecasting. The development and deployment of the pipeline are then discussed by Keerthisree.
Keerthisree talked about the development pipeline and steps involved in it such as data slicing, feature engineering, data cleaning, among others. She further discussed the overview of the deployment pipeline and the steps involved in it, which are functional testing, integration testing, acceptance testing, etc.
Lastly, the discussion was concluded by stressing on the future enhancements of the model that include model tweaking functionality, model monitoring, optimisation, logging enhancement and scheduler.