dataaaaa!

🎙️ It's About Data — Paul Blankley discusses the intersection of AI agents and data semantic layers. The discussion focuses on how agents interact with semantic layer architecture in modern data engineering stacks.

Omni features business intelligence, data modeling, and embedded analytics capabilities with context modeling. It integrates across data platforms like Snowflake, Databricks, and dbt.

Otrium deployed Omni to enable AI self-service over a governed semantic model, cutting ad hoc data tickets. Data engineers synced dbt docs and built curated datasets with joins and access controls, enabling accurate querying across LLM interfaces like Claude via an MCP server.

As Apache Ossie standardizes semantic models, differentiation shifts to query-time behavior. Agent workloads require dynamic SQL generation for ad-hoc transforms without config changes, plus hybrid search over entities and rules, moving beyond static caching and simple keyword lookups.

A context layer feeds AI agents unstructured business documentation alongside structured data. Unlike a semantic layer defined in dbt or LookML, it captures query history, joins, and internal notes from Slack or Notion to refine SQL accuracy and verify automated data transformations.

🎙️ Maxq Analytics — The open source agent nao queries data warehouses via natural language across Slack and Claude MCP. Because text-to-SQL is not enough, it integrates a context layer above the semantic layer to streamline consumption.

Graphene is an open-source analytics framework tailored for AI coding agents. It pairs an ANSI-compliant semantic layer with a code-based dashboard format to ensure accurate metrics and joins. Governed by a CLI, it brings version control, CI testing, and automated iteration to data pipelines.

OpenAI launched the Data agent in ChatGPT Work, connecting directly to Snowflake, Databricks, BigQuery, and Redshift. It integrates with dbt and semantic layers to enforce data models, while respecting existing row- and column-level access controls.

🎙️ It's About Data — Justin Borgman argues that building data products requires context engineering, emphasizing that context engineering is a human in the loop process. Data engineers must integrate domain knowledge directly into data architectures.

Cube Store is a distributed columnar engine written in Rust for serving pre-aggregations. Using Apache Arrow and DataFusion, it maintains indexes as sorted copies of Parquet data to skip non-matching rows and run streaming aggregations directly over partitioned worker caches.

Customer reporting requires robust data modeling, standardized metrics, and query-level multi-tenant isolation. Instead of relying on vulnerable UI filters, data teams must implement row-level security at the query layer via runtime identity tokens to prevent multi-tenant data exposure.

ChatGPT generates Python scripts via Code Interpreter, while Claude renders interactive artifacts via MCP connectors. Both suit prototyping, but data engineers cannot use them as production dashboards due to missing semantic layers, absent governance, lack of SLAs, and manual schema upkeep.

Embedded analytics adoption fails when users do not trust data or reach insights fast. Data engineers should implement a governed semantic layer and tenant-safe row-level security to prevent metric drift and cache issues. Track behavioral events rather than seat counts to measure real usage.

Claude builds interactive dashboards via MCP connectors to query live SQL databases and warehouses. While effective for prototyping, it lacks production data engineering requirements: there is no shared semantic layer, unified access control, or automated pipeline maintenance when schemas change.

Pointing AI agents at raw warehouse tables causes inconsistent metrics. To ensure reliability, teams must implement a semantic layer and tested data models so agents query validated business definitions instead of guessing schema nuances. Exploration can be probabilistic, but published BI cannot.

Motley integrates with GitHub to manage semantic models via Git workflows. The default branch serves as the source of truth, auto-importing configs and propagating updates. Teams can validate models inside PRs against data sources and test changes on isolated working branches.

me&u migrated four BI tools to Omni to unify their regionalized Snowflake data stack. By implementing YAML-based semantic modeling in Git alongside dbt, the team established governed joins and definitions that support direct warehouse querying, self-service dashboards, and Claude via MCP.

Centralizing data is not enough; retail and CPG pipelines require dbt to validate data integrity. Automated dbt tests catch broken feeds, while the dbt Semantic Layer defines shared metrics. Lineage graphs via dbt Explorer trace downstream numbers back to source systems for reliable automation.

Raw LLMs and AI tools fail at business scale due to lacking an enforced semantic layer, schema maintenance, and row-level permissions. Managed AI analytics services bridge this gap by building governed, operated data apps that ensure metric consistency, reliable pipelines, and audited access.

🎙️ The Joe Reis Show — Conceptual data modeling is vital for reliable AI, semantic layers, and LLMs. Bypassing data modeling creates architectural debt like 15,000 unmaintained Snowflake tables. Defining business concepts remains a prerequisite for sound architecture.

Benchmarking open-source Cube against SLayer failed when an agent solved zero tasks. Cube schema metadata was too verbose for context limits, and open-source Cube lacks APIs to dynamically add custom measures mid-task without building external model stores.

dst provides a governed data serving interface between warehouses and AI. Instead of generating unvalidated queries, it resolves plain-language requests using predefined team data definitions, returning SQL, a confidence grade, and execution receipts. It runs on Python and Docker with Postgres.

Building context for data AI agents requires structuring multiple operational layers: repo, pipeline, asset, semantic, and glossary. Integrating tools like CLI, MCP, and custom skills allows agents to execute actions, resolve layer conflicts, and maintain up-to-date context pipelines.

Data engineers must trace AI decisions across pipelines, not just models. In one case, an unflagged upstream schema change silently degraded a join in the semantic layer, corrupting feature data while pipelines ran green.

🎙️ Tony Zeljkovic — Deploying analytics agents requires more than connecting an LLM to a database. Nao co-founders discuss building open-source agentic analytics, delegating workflows, and shifting analytics engineers toward context engineering with semantic layers and metadata to replace traditional dashboards.