Case Study: Microsoft improves Journal AI models with Weights & Biases

How Microsoft Leveraged Weights & Biases to Build the Models Behind Ink

Microsoft, the customer, faced the challenge of developing the complex machine learning models required for its Journal app to intelligently interpret freeform ink gestures and note structures. The distributed nature of the team and the long training times for these AI models made collaboration difficult and created significant risks from using incorrect or untracked data versions.

The vendor, Weights & Biases, provided a solution through its Artifacts product, which acted as a "rock solid versioning system" for data. This gave the team a single source of truth, ensuring they were always using the correct data and understood its full lineage. The solution improved remote collaboration, saved valuable compute and iteration time by catching overfitting early, and provided a transparent, centralized platform for all experiments, which was a massive upgrade over their previous tools.

Microsoft

Dan Nissenbaum

Principle Research Software Development Engineer