Problem
Data analysts at Grab's Integrity Analytics team get a growing stream of requests from stakeholders. Most of them need the same SQL queries with small changes to dates, age groups or other filters. Running them by hand is slow and keeps analysts from higher-value work. Regular metric reports and fraud investigations are the heaviest cases.
Example from the article: the A* bot lets a fraud investigator send one prompt in Slack. It picks the relevant stored queries, runs them and returns a summary, which cut an investigation to minutes.
Solution in brief
Package the frequently used SQL queries as APIs on an internal platform called Data-Arks. An LLM application (SpellVault) uses RAG to pick the relevant API for the user's request, calls it, and writes a short summary of the returned table. Slack is the interface; a scheduler triggers routine reports. The LLM never writes SQL, so the data comes from vetted queries.
Questions this case answers
- How to answer recurring data questions with an LLM when the needed queries are already known and only their parameters change
- How to give an LLM access to fresh warehouse data without retraining it and without letting it generate unvetted SQL
- How to let an LLM answer recurring analytical questions without having it write SQL?
- How to automate regular metric reports and add a written summary of key insights?
Tools and methods
- Data-Arks — in-house Python API platform that exposes uploaded SQL queries and Python scripts as APIs
- SpellVault — in-house platform for storing and sharing LLM prompts, with low/no-code RAG
- RAG — selects the most relevant queries or plugins for a user request
- Tool / function calling — the LLM calls Data-Arks APIs and receives data
- Slack — user interface for commands and delivered reports
- Apache Airflow — scheduler that triggers the report app at set times
Google or a one-time email code. No password.