Retrofit a messy warehouse for AI agents: semantic layer, text-to-SQL, evaluation, governance, MCP and production ops
Your CEO just forwarded a conversational-analytics demo and said: “By Q3, anyone should ask our data in English and trust the answer.” That mandate lands on the data platform — not the app team. This course is that retrofit, built end to end on a real, messy, mid-migration warehouse.
What you’ll learn
- Diagnose exactly why naive LLM text-to-SQL hallucinates, and measure failure with a real evaluation harness..
- Design and author a semantic layer — metrics, dimensions, entities, and a join graph — that grounds an LLM and kills hallucinated joins..
- Lift text-to-SQL accuracy from roughly 40% to roughly 90% through grounding and context engineering for the warehouse..
- Choose a semantic-layer engine (dbt Semantic Layer, Cube, Snowflake Semantic Views, Databricks Metric Views, LookML) with a decision matrix and defend it..
- Enforce governance at the semantic layer so agents inherit RBAC, row/column security, and PII masking instead of bypassing them..
- Build and secure a governed MCP server that exposes Snowflake, dbt, and Databricks to agents safely..
- Operate agentic analytics in production with tracing, cost control, semantic-drift defense, and a rollout runbook..
- Apply agent patterns correctly — and recognize when NOT to use an agent at all..
Course Content
- The Mandate: Just Let Them Ask in English –> 4 lectures • 14min.
- Why Naive Text-to-SQL Hallucinates –> 4 lectures • 16min.
- What a Semantic Layer Actually Is –> 5 lectures • 20min.
- Modeling Metrics and Dimensions an LLM Can Trust –> 5 lectures • 24min.
- The Join Graph: Killing Hallucinated Joins –> 4 lectures • 21min.
- Choose Your Engine: The Semantic-Layer Decision Spine –> 4 lectures • 20min.
- Build It: A Working Semantic Layer –> 5 lectures • 21min.
- The Grounding Jump: 40 Percent to 90 Percent –> 5 lectures • 24min.
- Context Engineering for the Warehouse –> 5 lectures • 22min.
- How Do You Know It’s Right? Evaluation –> 4 lectures • 18min.
- The Governance Gap: What Breaks When Agents Query –> 4 lectures • 21min.
- Governance at the Semantic Layer –> 5 lectures • 26min.
- Guardrails: Containing the Runaway Agent –> 4 lectures • 20min.
- Why MCP: The Standard Contract Between Agents and Data –> 4 lectures • 18min.
- Build a Governed Warehouse MCP Server –> 5 lectures • 25min.
- MCP Architecture Decisions –> 4 lectures • 22min.
- Securing MCP –> 4 lectures • 21min.
- From One Query to Agentic Analytics –> 5 lectures • 22min.
- Agent Patterns (and When NOT to Use Them) –> 4 lectures • 21min.
- Observability: Trace the Wrong Answer Back –> 4 lectures • 20min.
- Production Hardening and the Drift Problem –> 4 lectures • 22min.
- Capstone: Ship and Defend an Agent-Queryable Semantic Layer –> 3 lectures • 16min.
- Introduction –> 2 lectures • 10min.
Requirements
Your CEO just forwarded a conversational-analytics demo and said: “By Q3, anyone should ask our data in English and trust the answer.” That mandate lands on the data platform — not the app team. This course is that retrofit, built end to end on a real, messy, mid-migration warehouse.
You will follow Maya, Head of Data Platform at a $2B retailer, as she makes an inherited Teradata-to-Snowflake estate safe for AI agents to query in plain English. You start by measuring why naive LLM text-to-SQL fails — hallucinated joins, three definitions of “revenue”, confident wrong numbers — using a real evaluation harness. Then you fix it.
You design and author a semantic layer (metrics, dimensions, entities, a join graph) that grounds text-to-SQL accuracy from roughly 40% to roughly 90%. You choose a semantic-layer engine with a decision matrix — dbt Semantic Layer, Cube, Snowflake Semantic Views, Databricks Metric Views, LookML — and learn to defend the choice.
From there you make it production-grade: enforce governance at the semantic layer so agents inherit RBAC, row/column security, and PII masking instead of bypassing them. You build and secure a governed MCP server that exposes Snowflake, dbt, and Databricks to agents safely. You add guardrails against the $40K runaway query, then operate agentic analytics with tracing, cost control, and semantic-drift defense.
By the capstone you ship an agent-queryable platform and defend every decision against an 8-dimension architect’s rubric. This is advanced, architect-level material for engineers who own the warehouse — not a prompt-engineering tour. You leave able to say “yes, safely” to the English-query mandate.