Grab’s Analytics Data Warehouse team faced a bottleneck: engineers spent two days answering data questions. Managing over 15,000 tables that serve half the queries in the company’s data lake, the team fielded 1,000 monthly users. The shift also aimed to reduce silos and improve visibility as AI‑driven features accelerate release cycles. To free up expertise, they built an AI system that automates the investigation workflow.

The architecture separates the reasoning LLM—dubbed the “brain”—from specialist “hands” that fetch metadata, run SQL, and query internal platforms. FastAPI receives Slack requests, LangGraph orchestrates looping agent interactions, Redis caches sessions and PostgreSQL logs conversation history. Agents pull from Hubble (catalog), Genchi (data quality) and Lighthouse (pipeline health), enabling precise, auditable answers. The design also logs each step for post‑mortem analysis.

Two pathways handle read‑only investigations and write‑back enhancements. A classifier routes queries to four agents—Data, Code Search, On‑call and Summarizer—while a single Enhancement agent drafts schema changes and opens merge requests, always awaiting human review. By modularizing agents, Grab tolerates latency spikes but gains maintainability, turning hours of manual tracing into minutes of reliable AI‑assisted responses.