TechPrufer
Get started
Grab
Grab logoGrab · Feature trail

AI-Assisted Analytics

11 posts · Jul 2019 – Jul 2026

What this trail covers

Grab's AI-Assisted Analytics initiative has evolved with Metasense V2, an LLM-powered system for data governance. This iteration enhances data classification by splitting the model into PII and non-PII taggers, refining prompts for better accuracy with complex data structures, and improving model capacity by reducing tag count and handling large tables. The integration of LangChain and LangSmith frameworks streamlines prompt experimentation, evaluation, and deployment, enabling rapid iteration. The Gemini orchestration service now integrates LLM classification for column-level tagging, addressing challenges with third-party tools and manual classification. It manages requests, aggregates them into batches, and employs rate limiting. The LLM is prompted to act as a column tag classifier, assigning tags based on table and column names, with a defined set of tags and a JSON output format. Techniques like clear articulation of requirements and few-shot learning are used for prompt engineering.

All Grab trails
  1. Grab Bench: Evaluating AI on Grab-shaped production workThis post introduces Grab Bench, a new evaluation harness for AI systems designed to assess AI models on Grab-shaped work. It focuses on evaluating specific contracts such as metric faithfulness, tool-parameter discipline, evidence grounding, safety boundaries, and repository-level behavior. Grab Bench utilizes task plugins for various AI tasks including SQL generation, tool use, profile reasoning, and agentic coding, employing deterministic scorers or LLM judges. The system emphasizes safe case design using synthetic or redacted data, scoring contracts over confidence, and making shortcuts visible through baselines and failure taxonomies. It supports internal reproducibility with teaching and certification artifacts and aims to evolve towards pipeline-like evaluations and easier reuse by other teams.
  2. How AI is transforming analytics at GrabThis post details the evolution of Grab's AI-assisted analytics capabilities, moving towards more autonomous workflows. It introduces Spartan, an end-to-end agentic analytics workflow accessible via Slack, which uses over 50 skills and 120 analysis frameworks to answer business questions in natural language. It also describes Scarlet, an agentic system for near-self-healing pipelines that automates root-cause analysis and triage. The post highlights the importance of context management with ContextIQ and its Context Lifecycle Manager, enabling agents to maintain and update knowledge bases. It also introduces BriX, an internal portal built by data engineers to create purpose-built analytics surfaces, democratizing L3 workflows for non-analysts. The post discusses the 'ladder' of AI autonomy in analytics, from AI-Assisted (L2) to End-to-end autonomous (L5), emphasizing the role of human judgment in problem framing, metric definition, and decision-making.
  3. Palana (Part 1): Why Grab built a secure platform for autonomous AI AgentsThis post introduces Palana, Grab's Kubernetes-native platform for running autonomous AI agents securely. It details the platform's design principles including isolation as the unit of trust, never giving credentials directly to the agent, making egress a control point via Envoy and OPA, and ensuring the control plane remains outside the agent. It also highlights the use of Kubernetes primitives for managing agent lifecycles and resources.
  4. Palana (Part 2): Architecting isolation, identity, and auditability for AI agentsThis post dives into the technical architecture of Palana, Grab's secure execution platform for AI agents. It details the implementation of isolation, identity, and auditability through Kubernetes namespaces, RBAC, network policies (Cilium), and proxy-based access control (Traefik, OAuth2-Proxy, OPA). It also explains the secrets management strategy using Vault, LLM access via a dedicated proxy wrapper, and the observability mechanisms including structured logging and an idle-shutdown reaper. The post highlights lessons learned regarding platform-level security, user experience, credential handling, namespace compounding, and lifecycle management for long-running agents.
  5. Agent platform (Part 1): How we help Grab build and run AI agents at scaleThis post details the initial architecture of Grab's AI support bot, which served as the precursor to LLM-Kit. It describes the reasoning plane (single-agent loop) and the tool plane (integrating Glean, Kibana, GitLab, Slack). It also outlines the key pain points encountered during scaling: lack of robust evaluation strategies, the need for fast model and provider switching, the critical importance of observability, and the significant effort required for production plumbing (auth, secrets, config, etc.). These pain points directly led to the development of the LLM-Kit framework, which aims to standardize these surrounding components.
  7. From firefighting to building: How AI agents restored our team’s core productivityThis post details the development and deployment of a multi-agent AI system to automate responses to data-related questions and assist with data pipeline enhancements for Grab's Analytics Data Warehouse (ADW) team. The system uses FastAPI and LangGraph for orchestration, with specialized agents (Data Agent, Code Search Agent, On-call Agent, Enhancement Agent) interacting with tools like Trino, Hive, Delta Lake, GitLab, Apache Spark, Airflow, Confluence, and Slack. It addresses the challenge of repetitive requests by automating context hunting and code changes, improving engineering productivity. The post contrasts this with a monolithic AI approach, highlighting the benefits of modularity and focused expertise. It also outlines two key workflows: semi-automated code changes for enhancement requests and a multi-agent investigation pathway for data issues.
  8. How we built a custom vision LLM to improve document processing at GrabDeveloped a custom lightweight Vision LLM (~1B parameters) from scratch by combining the vision encoder from Qwen2-VL 2B and the language decoder from Qwen2.5 0.5B, connected by an adjusted projector layer. Implemented a four-stage training process: projector alignment, vision tower enhancement on public multimodal datasets, language-specific visual training with synthetic OCR data, and task-centric fine-tuning. Achieved performance comparable to the larger 2B model with significantly reduced latency, outperforming traditional OCR and external APIs.
  9. User foundation models for GrabThis post introduces Grab's development of a user foundation model, a transformer-based architecture that learns from both tabular and time-series user interaction data. It addresses the challenge of integrating diverse data modalities (text, numerical, IDs, locations) and scaling to millions of entities. The model generates user embeddings that generalize across various downstream tasks like recommendation, advertisement optimization, fraud detection, and churn prediction, moving away from manually engineered features. The architecture unifies order-invariant tabular data with order-dependent time-series data using a `key:value` tokenization strategy.
  10. Metasense V2: Enhancing, improving and productionisation of LLM powered data governanceMetasense V2 enhances Grab's LLM-powered data governance by splitting the classification model into PII and non-PII taggers, reducing the number of non-PII tags from 21 to 8, and splitting tables with over 150 columns. It integrates LangChain for prompt chaining and LangSmith for collaborative prompt experimentation, evaluation, and monitoring. This leads to streamlined prompt optimization, transparent performance metrics, and a protocol for continuous improvement based on misclassification alerts.
  11. LLM-assisted vector similarity searchThis post introduces a novel approach to enhance search relevance by combining vector similarity search with LLMs. The proposed method involves an initial vector similarity search to shortlist potential results, followed by an LLM that ranks these shortlisted results based on contextual relevance. Experiments on synthetic datasets showed that this LLM-assisted approach significantly outperforms raw vector similarity search for complex queries involving negations or conceptual requirements, as demonstrated by examples from food and tourist spot datasets. The post also notes that for simpler queries, both methods perform comparably.
  12. Leveraging RAG-powered LLMs for analytical tasksIntroduced Data-Arks, an in-house Python-based API platform that hosts frequently used SQL queries and Python functions, packaging them into individual APIs. Data-Arks integrates with Slack, Wiki, and JIRA APIs to parse and fetch information. This platform serves as middleware to channel data to LLM agents, enabling automation of regular metric reports and fraud investigations. The A* bot, an LLM fraud investigation helper, leverages Data-Arks and Spellvault to execute relevant fraud investigation queries and provide summarized results via Slack. The post also elaborates on the rationale for choosing RAG over fine-tuning for LLM accuracy improvements, citing effort/cost, ability to respond with latest information, and speed/scalability.
  13. Enabling conversational data discovery with LLMs at GrabThis post details the evolution of Grab's data discovery platform, Hubble, by integrating LLMs to create HubbleIQ. The technical contributions include: 1. Enhancing Elasticsearch for exact, partial, and inexact search queries by tagging P80 tables, boosting relevant schemas, hiding irrelevant datasets, deboosting deprecated tables, improving the search UI, adding tags, and boosting certified tables, resulting in a 12 percentage point increase in click-through rate to 94%. 2. Building a context store for HubbleIQ by developing a GPT-4 based documentation generation engine that uses table schemas and sample data, increasing documentation coverage for P80 tables by 70 percentage points to ~90%. 3. Launching HubbleIQ by integrating Hubble with Glean, creating a bot using Glean Apps with a custom system prompt, and integrating it into Hubble search and Slack. This has reduced data discovery time and improved user satisfaction.
  14. LLM-powered data classification for data entities at scaleThis post details the evolution of Grab's data classification system, Gemini, to incorporate LLM-powered column-level tagging. It describes the challenges with previous approaches (manual classification, third-party tools with false positives) and how LLMs address these by providing a natural language interface for data governance personnel. The methodology includes an orchestration service that manages requests, aggregates them into batches, and applies rate limiting. The classification process focuses on LLM-powered column-level tag classification, defining the task as assigning tags to schema fields based on an internal classification scheme. A detailed prompt example for the LLM is provided, along with a curated tag library. Prompt engineering techniques like articulating requirements and few-shot learning are highlighted as crucial for achieving a fixed output format.
  15. Scaling marketing for merchants with targeted and intelligent promosThis post introduces Bullseye, an AI-driven promo assignment system designed to optimize marketing campaigns for merchants. Bullseye addresses the limitations of globally assigned promos by implementing automated user segmentation, automatic promo design tailored to individual merchant goals, and a customer response model that predicts user behavior based on attributes. The system uses a Spark job for prediction and promo assignment, with components like Eater Segments Identifier, Promo Designer (including Offer Content Candidate Generation, Campaign Impact Simulator, and Optimal Promo Selection), and Customer Response Model. Hyper-parameter tuning is used to optimize for multiple business goals, balancing sales uplift with promo spending efficiency.
  16. PII masking for privacy-grade machine learningThis post details the implementation of PII masking for machine learning data streaming pipelines at Grab. It describes a process involving PII tagging in protobuf schemas using a custom `pii.proto` library, a CI pipeline with a Python script to enforce tagging and a whitelist for exceptions, and a Flink application that consumes production data, dynamically masks PII based on schema tags (using HMAC for strings and seeded random generators for numbers/dates), and publishes sanitized data to the staging environment. The goal is to enable data engineers to tune ML models with production data while maintaining privacy by design.
  17. Automated Experiment Analysis - Making experimental analysis scalableThis post introduces the Automated Experiment Analysis platform, a data product that standardizes and automates the statistical analysis of experiments (A/B tests). It addresses inefficiencies and quality control issues of manual analysis by implementing a metrics glossary, automated data pipelines (bronze, silver, gold datasets), and a "Decision Engine" for statistical analysis. Results are visualized via embedded Power BI. Key technical components include Azure Data Lake for metadata storage, Spark for ETL jobs on Databricks, a Python package "Decision Engine" for statistical tests, and embedded Power BI for visualization. Future plans include a centralized metric store, end-to-end experiment guide rails, and a more comprehensive analysis toolbox.
  18. Making Grab’s Everyday App SuperThis post details the evolution of the Grab Feed recommendation engine, introducing strategies such as popularity, user favorites, collaborative filtering, habitual behavior, and deep recommendations that leverage cross-business data. It also discusses the use of an experimentation platform to measure recommendation quality through online and offline metrics, and outlines future work including multi-objective optimization, hyperparameter optimization automation, online learning, multi-armed bandits, and a recsplanation system.