Skip to content
Yash Singh Ramgadiya

Case study

Organisation
Burda Luxury, Asia Pacific

In-house AI-enabled Analytics Agent

In-house conversational analytics agent combining governed semantic querying, tool orchestration and production analytical workflows.

What changed

Natural-language analytics moved onto governed semantic APIs rather than bypassing the metric layer.

Agentic AnalyticsSemantic LayerAnalytics Product
How the AI-enabled analytics agent answers a questionA business question is handled by agent orchestration, which inspects the semantic catalogue, runs a governed semantic query, validates the result with retries where needed, and returns an insight, chart or dashboard artefact. Responses stream back to the user throughout.Business questionAgent orchestrationSemantic catalogueGoverned semantic queryValidation and retryInsight, chart or dashboard artefactStreamingresponse
How the AI-enabled analytics agent answers a question. Model routing sits inside orchestration; every answer is anchored on governed semantic APIs.

Context

Leaders and operators across 8 brands and 5 APAC markets needed answers faster than a ticket-driven analytics queue could provide, without giving up the governance that makes numbers trustworthy.

The organisation already had a governed semantic layer. The opportunity was to put a conversational interface on top of it rather than beside it.

Challenge

Most conversational-analytics attempts fail on trust: free-form SQL generation drifts from governed definitions, answers are hard to validate, and failures are silent.

The agent had to answer in production, on governed metrics, with visible validation, sensible retries and responses that stream rather than stall.

My Role

I lead the programme as part of the in-house production analytics platform, setting the product direction, the governance boundaries the agent must respect and the engineering priorities of the cross-functional team delivering it.

Approach

  • Anchor every answer on governed Cube semantic APIs so the agent queries defined measures and dimensions rather than raw tables.
  • Give the agent semantic-catalog inspection so it can discover what is measurable before it attempts a query.
  • Model the workflow as tool orchestration: catalog inspection, governed query execution, chart generation and dashboard-artefact generation.
  • Validate outputs and handle retries before anything reaches the user.
  • Stream responses so users see progress and partial results instead of waiting for a single final payload.

Architecture / Analytics Design

The agent is orchestrated with LangGraph and LangChain, with LiteLLM providing model routing across providers. Tool orchestration connects the orchestration layer to governed Cube semantic APIs, chart generation and dashboard artefacts. Validation and retry handling wrap query execution, and responses stream back to the interface.

The design deliberately keeps business logic in the semantic layer, not in prompts, so that definitions stay governed and auditable.

Business Impact

The agent is a production conversational analytics capability that lets teams interrogate governed metrics, generate charts and produce dashboard artefacts within business workflows, reducing dependence on ad hoc analyst requests.

Technology

  • LangGraph
  • LangChain
  • LiteLLM (model routing)
  • Cube semantic APIs
  • Semantic-catalog inspection and tool orchestration
  • Validation and retry handling
  • Streaming responses

Key Lessons

  • Governed semantics are the prerequisite for trustworthy AI analytics; a great model on ungoverned data still produces ungoverned answers.
  • Validation and retries are product features, not implementation details.
  • Streaming changes how people perceive latency and how willing they are to iterate.

Related work