SeenPixelReference design · not client work
A governed ask-your-data agent
The architecture a pilot starts from. Each component is something we build or configure inside your cloud account; the warehouse is the one you already have. A live walkthrough is available on request.
Components
User or assistant
A person in Claude, ChatGPT, Slack or an IDE. They ask in plain language; they are authenticated through your SSO before anything else happens.
Agent runtime
Plans the answer, calls tools, explains what it did and cites the metric definitions it used. Model-agnostic; swapping the model does not touch governance.
MCP server
Exposes a small set of typed tools (list metrics, describe a metric, run a metric query) with schemas and tests. The agent cannot run free-form SQL.
Auth and policy
Maps the user to warehouse roles, applies row- and column-level rules, writes an audit record for every call. Runs as the user, never as a service account with everything.
Semantic layer
Cube, dbt metrics or LookML: the one place where 'active customer' and 'net revenue' are defined. Compiles metric queries to governed SQL.
Warehouse
Your existing Snowflake, BigQuery or Databricks. Nothing is copied out; the warehouse's own permissions are the last line.
Evaluations and monitoring
Golden questions with known answers run on every change; accuracy, latency and cost per query tracked; traces kept; alerts when any of them move.
What the design is answering
| Typical failure | In this design |
|---|---|
| Agent writes plausible SQL over raw tables; numbers disagree with finance. | Metric queries only, through the semantic layer. |
| A user sees rows they should not. | Policy enforced below the agent and in the warehouse, running as the user. |
| It worked in the demo; nobody knows if it still does. | Evaluation suite in CI, with accuracy per question over time. |
| The bill for one chatty user surprises everyone. | Cost per query measured, with ceilings and alerts. |
Why this matters: Gartner expects 60% of agentic-analytics projects relying solely on MCP to fail by 2028 without a consistent semantic layer (Gartner, May 2026), and only 11–14% of MCP pilots reach production (Stacklok, State of MCP in Software 2026). The design exists to be in the other group.
Live demo on request
There is no public demo link. We walk through this design live, in a 30-minute session, on request. Ask through the contact form and mention “ask-your-data demo”.
Want the walkthrough?
A 30-minute live session on this design, then a conversation about what it would take over your warehouse.