Data agents that check their own numbers

Business teams could not query the warehouse themselves, so every KPI question became a ticket for an analyst.

Pointing a language model at a raw warehouse is worse than no answer: it invents tables that do not exist and numbers nobody can trace.

A multi-agent system on Vertex AI Agent Engine, built with Google ADK. A semantic router sends each question to the right specialist agent. SQL generation is M-Schema aware, so the agent works from the real schema instead of guessing it.

A grounding guardrail validates every number against live BigQuery results before it reaches a user. Multi-turn memory, query planning and anomaly detection sit around it, and when confidence is low the agent abstains instead of guessing. Answers come back as a short narrative with a chart and a table.

A sibling agent turns natural language into DAX for Power BI. To test it I built an adversarial QA harness: probes that feed the agent inputs the test suite never covered, score the answers, and turn every failure into a permanent regression case.

1. questionplain language, multi-turn
2. semantic routerpicks the specialist
3. SQL agentM-Schema aware
4. grounding guardrailchecked against live BigQuery
5. answer, or abstainno unverified numbers

In production at Maxis, answering telecom KPI questions for non-technical users. The grounding loop eliminated hallucinated figures: every number shown has been checked against the warehouse that produced it.

I own the LLMOps loop end to end: containerised Cloud Run deploys, versioned prompts and agent configs, reliability and cost monitoring.

Internal system, so no screenshots or data are shown here.

• Python
• Google ADK
• Vertex AI Agent Engine
• BigQuery
• Cloud Run
• Power BI / DAX