Marketing ops · Production

Grab builds a user foundation model generating embeddings for personalisation across its superapp ecosystem

The problem

Grab's recommender systems relied on hundreds to thousands of manually engineered, task-specific, siloed features that required substantial effort and could not effectively capture sequential interaction data. General-purpose LLMs lacked the contextual understanding for Grab's domain-specific data, and off-the-shelf models could not jointly handle the superapp's mix of tabular, sequential, and multi-modal data.

First attempt

General-purpose LLMs lacked the contextual understanding required for Grab's domain-specific data, and single-task supervised training would produce biased embeddings unsuitable for Grab's diverse verticals.

Workflow diagram · grounded in source
1
User interaction data collection
trigger
“Grab tracks every interaction on its app, including what users view, click, consider, and ultimately transact”
2
Key-value tokenisation
ai_action
“We define a universal token structure as a key:value pair. - For tabular data, the key is the column name (e.g. online_hours ) and the value is the user's attribute (e.g. 4 ). - For time-series data, the key is the event type (e.g. view_…”
3
Unsupervised foundation model pre-training
ai_action
“To pre-train our model on tabular and time-series data, we combine masked language modeling (reconstructing randomly masked tokens) with next action prediction.”
4
Daily batch embedding inference
ai_action
“To generate fresh embeddings for millions of users, we leverage Ray Data—an open-source library used for data processing in AI and Machine Learning (ML) workload, to execute a distributed batch inference pipeline.”
5
Long-term and short-term embedding extraction
output
“Our architecture is deliberately designed to produce two distinct but complementary types of user embeddings, providing a holistic view by capturing both the user's stable, long-term identity and their dynamic, short-term intent.”
6
Downstream model serving
integration
“These embeddings serve as rich, general-purpose features that can support a wide range of separate downstream models.”
Reported outcome

Grab's foundation model now powers ad optimisation, dual app prediction, fraud detection, and churn probability; the distributed Ray infrastructure dramatically reduces costs and accelerates processing times; and teams building on pre-trained embeddings see significantly reduced development time and improved performance.

Reported metrics
Manually engineered features (legacy baseline)hundreds to thousands
Infrastructure costdramatically reduces costs
Processing timeaccelerates processing times
Downstream model development timesignificantly reducing development time
Show all 5 reported metrics
manually engineered features (legacy baseline)hundreds to thousands
infrastructure costdramatically reduces costs
processing timeaccelerates processing times
downstream model development timesignificantly reducing development time
downstream model performanceimproving performance
Reported stack
RayRay Data
Source
https://engineering.grab.com/user-foundation-models-for-grab
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Grab's foundation model now powers ad optimisation, dual app prediction, fraud detection, and churn probability; the distributed Ray infrastructure dramatically reduces costs and accelerates processing times; and team…

What tools did this team use?

Ray, Ray Data.

What results were reported?

Manually engineered features (legacy baseline): hundreds to thousands; Infrastructure cost: dramatically reduces costs; Processing time: accelerates processing times; Downstream model development time: significantly reducing development time (source-reported, not independently verified).

What failed first in this deployment?

General-purpose LLMs lacked the contextual understanding required for Grab's domain-specific data, and single-task supervised training would produce biased embeddings unsuitable for Grab's diverse verticals.

How is this marketing ops AI workflow structured?

User interaction data collection → Key-value tokenisation → Unsupervised foundation model pre-training → Daily batch embedding inference → Long-term and short-term embedding extraction → Downstream model serving.

Related marketing ops cases