Post by ByteByteGo

🌐 How Grab Built an AI Foundational Model to Understand Customers Better Grab is no longer just a ride-hailing app. It is a superapp with food delivery, groceries, mobility, payments, and financial services. This creates an enormous stream of user interactions, but until recently, personalization relied heavily on manually engineered features that were brittle and slow to scale. Grab's engineering team changed it by building a foundation model that learns directly from user data to make sense of long-term user traits and short-term intent. The model handles text queries, numerical values, and geolocation data to support a personalization engine that powers ads, recommendations, fraud systems, and other retention models across the company. A few key points that stand out: 1 - Grab designed a transformer-based architecture that uses a key-value token format and modality-specific adapters. 2 - Training the model involves unsupervised pre-training with masked token prediction and next action prediction. 3 - Model supports two usage modes. Teams can fine-tune it for specific tasks like fraud detection, or they can extract embeddings as universal behavioral features for other models. This model is helping Grab build a scalable AI foundation for how superapps will understand customers better. By learning directly from user data instead of relying on brittle manual features, the model gives Grab a scalable foundation for personalization across the superapp. 🔗 Read the full breakdown: https://lnkd.in/eD_fguVB Supported by our partners building tools that help engineering teams maintain reliability as personalization systems grow in complexity: Sentry - AI Code Review that catches issues early and improves code quality as your codebase and models scale. ➡️ https://bit.ly/3M8t3SE