Agentic data discovery for a telecom operator
Plain-language access to the data warehouse, no SQL needed
- Sector
- Telecommunications
- Market
- Southeast Asia
82.3%answer accuracy, above the 80% target
91.4%clarity score
2languages, English and Bahasa Indonesia
The challenge
Business users needed SQL to get insight, technical data definitions didn't match business KPIs, expert knowledge was siloed, and there was no automated lineage to show the impact of upstream changes.
What was built
- A stateful ReAct agent on LangGraph: verify the input, think, act, answer
- Business-glossary lookups from internal documentation
- Plain-language discovery of warehouse tables and data marts
- Live warehouse queries through a Snowflake MCP server
How it was built
- Qwen 3-32B on Amazon Bedrock for intent analysis and SQL generation
- A Bedrock knowledge base on OpenSearch Serverless for retrieval