DataClaw BI analysis agent

Make self-service analysis trustworthy enough to ship.

The platform replaced an external BI product and a tightly coupled ClickHouse query path. Analysts needed governed dashboards and ad-hoc answers, while the data estate moved across Hive, Presto, Doris and other OLAP engines. Metric definitions had to stay consistent as natural-language access arrived.

We delivered the BI product, multi-engine query layer and DataClaw agent workflow: MetricFlow semantic metrics, metadata grounding, SQL linting and review, MCP session pooling, and a tool loop that executes before it explains or charts.

In the work

BI replacement from zero
Dashboards, charts, spaces and permissions became a self-owned product instead of a vendor-bound surface.
Semantic grounding
MetricFlow definitions and metadata give NL2SQL a governed vocabulary across engines and dashboards.
SQL before presentation
Generate, lint, review, fix and execute are traced as one loop before a result is rendered or charted.

The problem

A mainland technology company needed to replace an external BI dependency while supporting 3,000+ dashboards, 20,000+ charts and multiple query engines. Self-service analysis was slow, and missing metric definitions made first-pass NL2SQL unreliable.

Measured result

The platform now carries 3,000+ dashboards and 20,000+ charts; semantic-cache hit rate rose from 50% to 90%+. Ad-hoc data delivery moved from days to the same day. DataClaw first-pass NL2SQL success is about 50%, while SQL governance blocks 80%+ of erroneous SQL before release.

Scope

BI product replacement: dashboards, charts, spaces and permissions
Presto + cloud Doris query architecture with semantic caching
DataClaw agent layer, MetricFlow semantic metrics and SQL governance

What shipped

→ Self-built BI platform from zero with multi-engine query execution
→ DataClaw Skills engine, LLM intent routing, Hive metadata API and MCP session pool
→ SQL Linter → LLM review → repair loop with execute-before-chart guardrail

Anonymisation

The client, product names and commercial data are anonymised. Platform scale, cache, delivery-time and SQL-governance metrics come from that engagement’s records.

Work

Anonymised delivery

HSK 3 learning agent, from zero

Anonymised delivery

Agent for P0 scheduling and data sync

Anonymised delivery

Data-platform Agent runtime and MCP tool plane

Anonymised delivery

Governed knowledge gate for analysis Agents

Anonymised delivery

Alert-to-remediation operations Agent

Anonymised delivery

Agent control plane with guarded actions

Start a conversation