Scale and stability

The large dataset generated by Telkomsel’s huge customer base makes it possible to build highly accurate AI models, but it needs massive, stable, scalable compute power to process 20-30 terabytes of data about customers’ behavior daily. ISYANA was the operator’s first experience of building such a large-scale platform, and was made more difficult by its shortage of engineers with the right skills, hence it followed guidance from TM Forum (see below).

Telkomsel adopted some well-known AI models like collaborative-filtering and matrix-factorization, and built some to meet business goals, like preventing churn, or increasing customers’ spending.

The operator pulls data related to customers behavior from across its data ecosystem, runs it through a cluster processing engine to aggregate the data and create a master table which speeds up data scientists’ development of AI models. This workflow prevents siloes between departments and ensures the data’s quality.

In the early days of building ISYANA, Telkomsel struggled to optimize its machine learning pipeline due to the volume of data and so it built a two-part DevOps framework to incorporate the recommendation system’s algorithm and deploy it at scale.

A background job element refreshes the recommendation periodically and aligns with the cluster-processing engine to process large amount of data. The other part retrieves recommendations from Telkomsel’s customer-facing applications and channels to present personalized recommendations every time a customer opens an application. /p>