TechPrufer
Get started
Grab
Grab logoGrab · Feature trail

AI-Assisted Development Tooling

9 posts · Apr 2021 – Mar 2026

What this trail covers

Grab's AI Gateway provides a centralized, unified interface for accessing multiple GenAI providers (OpenAI, Azure, AWS Bedrock, Google VertexAI, etc.). It streamlines access, enables experimentation with various LLMs, optimizes cost-efficiency through shared capacity, and enforces auditing and platform benefits like monitoring, cost attribution, and rate limiting. The architecture acts as a reverse proxy, handling authentication, authorization, and rate limiting, while also supporting div. The AI tool Cursor has been adopted widely, with 98% of Tech Grabbers becoming monthly active users and 75% using it weekly. Cursor's suggestion acceptance rate is around 50%, indicating relevant and critically reviewed suggestions. Engineers utilize Cursor for unit test generation, code refactoring, cross-repository navigation, bug fixing, and automation. Custom tooling was built for monorepo indexing and preconfigured rules to align with Grab-specific coding conventions. Beyond engineering, AI-assisted development is expanding to non-technical teams and product design for direct production fixes. Productivity gains are observed to scale with usage intensity.

All Grab trails
  1. From firefighting to building: How AI agents restored our team’s core productivityThis post introduces a multi-agent AI system designed to automate responses to common data-related questions and enhancement requests within Grab's Analytics Data Warehouse (ADW) team. It details the architecture using FastAPI and LangGraph, with specialized agents (Data, Code Search, On-call, Enhancement) and supporting tools like Hubble, Genchi, and Lighthouse. The system handles two main pathways: enhancement requests (semi-automated code changes) and investigation questions (coordinated agent workflows). Key technical contributions include the agent orchestration logic, data investigation capabilities, code tracing, production system monitoring integration, and response summarization.
  2. Enabling R8 optimization at scale with AI-assisted debuggingThis post introduces the application of AI-assisted debugging for R8 optimization challenges in the Grab Android app. It details the development of custom MCP (Model Context Protocol) tools for automatic APK decompilation and stack trace deobfuscation, and their integration with GitLab CI for parallel testing of multiple proposed solutions. This approach significantly accelerates the debugging and resolution of R8-related issues by automating reverse engineering tasks and enabling rapid iteration through parallel CI builds.
  3. Cursor at Grab: Adoption and impactThis post details the adoption and impact of the AI coding assistant Cursor at Grab. It highlights a multi-tool strategy for AI coding assistants, with Cursor achieving 98% monthly active users and a 50% suggestion acceptance rate. Key utilization patterns include unit test generation, code refactoring, cross-repository navigation, and bug fixing. Custom tooling was developed for monorepo indexing and Grab-specific coding conventions. The post also touches on the expansion of AI-assisted development to non-engineering roles and the observed productivity gains.
  4. From deployment slop to production reality: How BriX bridges the gap with enterprise-grade AI infrastructureThis post introduces BriX, a platform designed to operationalize AI prototypes into enterprise-grade production systems. BriX addresses the challenges of 'deployment slop' by providing a production layer that handles security, scaling, data connections, and consistent behavior. Key capabilities include model agnosticism (switching between models like Claude or GPT), secure enterprise data connections via Model Context Protocols (MCPs), and centralized configuration of system prompts and context files to ensure consistent behavior. The BriX architecture is described as a synchronous streaming architecture with six layers: Frontend (React/TypeScript with SSE), Gateway (FastAPI), Brain (LangGraph orchestration), Memory (Hot/Cold Storage), Security (Identity Propagation), and Data Processing. The workflow involves uploading a prototype, configuring it through a non-coding interface, locking configurations, and deploying it. This enables domain experts to deploy AI tools in days instead of months.
  5. From failure to success: The birth of GrabGPT, Grab’s internal ChatGPTThis post details the origin and rapid adoption of GrabGPT, an internal ChatGPT-like tool. It highlights the pivot from a failed attempt at a platform support chatbot to building a company-wide AI assistant. Key technical contributions include leveraging existing frameworks, integrating Google login for authentication, deploying on Grab's private model-serving platform (Catwalk), and supporting multiple LLM providers (OpenAI, Claude, Gemini) on a private route for data security and global accessibility. The post emphasizes the success factors beyond technology, such as timing, security, and accessibility, and its role in catalyzing Grab's broader LLM strategy.
  7. Grab AI Gateway: Connecting Grabbers to multiple GenAI providersThis post introduces the Grab AI Gateway, a system designed to provide unified and streamlined access to multiple Generative AI (GenAI) providers (OpenAI, Azure, AWS Bedrock, Google VertexAI, etc.) for Grabbers. It details the architectural design as a set of reverse proxies, the IAM implementation using API keys with request path-based authorization, and the user journey including onboarding, exploration keys, a unified API interface (based on OpenAI's scheme), dynamic routing for usage efficiency, auditing of calls, cost attribution, and rate limiting. It also highlights integration with Grab's ML platform (Chimera notebooks, Catwalk). Challenges include keeping up with rapid innovations, fair distribution of quota between batch and online use cases, and maintaining reverse proxies.
  8. Supercharging LLM application development with LLM-KitIntroduces the LLM-Kit, a comprehensive framework for building production-ready Generative AI applications. It addresses challenges in security, scalability, integration, monitoring, and standardization. Key features include a pre-configured structure, integrated tech stack (Poetry, Gunicorn, FastAPI, LangChain, LangSmith, Vault, EKS, Gitlab CI), built-in observability (Datadog, LangSmith), efficient config/secret management, OIDC authentication, API documentation, Redis & vector database integration, a deployment pipeline, and evaluation frameworks. The post details the automated process of generating LLM applications, including code scaffolding, infrastructure provisioning (Terraform, ECR, EKS), and deployment. It also highlights the RAG scaffolding process for setting up PGVector databases.
  9. Leveraging RAG-powered LLMs for analytical tasksIntroduced Data-Arks, a Python-based API platform that hosts frequently used SQL queries and Python functions, allowing them to be called as APIs to serve data to LLM agents. This platform integrates with Slack, Wiki, and JIRA. Developed an LLM-powered fraud investigation helper (A* bot) and a Report Summarizer that uses Data-Arks APIs and LLMs to automate report generation and summarization, saving an estimated 3-4 hours per report. Detailed the rationale for choosing RAG over fine-tuning, emphasizing cost-efficiency, ability to respond with the latest information, and speed/scalability.
  10. How We Improved Agent Chat Efficiency with Machine LearningThis post details the development and implementation of SmartChat, a machine learning model designed to improve agent chat efficiency by providing contextual sentence completion suggestions. It covers the problem discovery, design iterations (including UI/UX considerations inspired by Gmail's Smart Compose), key metrics for tracking progress (characters saved, model effectiveness, acceptance rate, latency), the system architecture involving the Catwalk ML serving platform and API gateway, and the design of the ML model itself. The model uses a GRU-based Encoder-Decoder architecture with specific features like past conversations, time of day, and user type as inputs. The UI implementation uses a typeahead solution in ReactJS.