SeenPixelService
Governed “ask your data” agents over the warehouse you already have
A semantic layer, an MCP server, an evaluation loop and access control, so people can ask questions in plain language and get the same number the dashboard gives. Built in a 6–8-week pilot, in your environment.
What it is
Four parts that have to exist together. Leave one out and the project joins the statistics below.
Semantic layer
One definition of every metric and dimension, in Cube, dbt metrics or LookML, so the agent cannot invent a revenue number that disagrees with the dashboard.
Agents and MCP server
An MCP server that exposes typed tools over the semantic layer, and an agent runtime that plans, calls them and explains its answer. Works from Claude, ChatGPT, Slack or an IDE.
Evaluations
A set of golden questions with known answers, run on every change. Accuracy, latency and cost per query are measured, not assumed.
Access control
SSO, row- and column-level rules enforced below the agent, and an audit log of every query. A user gets the same answer the warehouse would give them directly.
Why most of these projects fail
60%
of agentic-analytics projects relying solely on MCP will fail by 2028 without a consistent semantic layer.
Source: Gartner, May 2026
11–14%
of MCP pilots reach production.
Source: Stacklok, State of MCP in Software 2026
71%
of data practitioners worry about hallucinated numbers reaching stakeholders.
Source: dbt Labs, State of Analytics Engineering 2026
The pattern is the same each time: an agent is pointed at raw tables, it writes plausible SQL, the numbers do not match finance, and trust is gone. The fix is not a better prompt. It is a layer that defines the metrics, a policy layer that enforces who sees what, and a test suite that proves the answers.
Reference architecture
The design a pilot starts from. Every component sits inside your cloud account.
What we deliver
- Semantic model for one or two business domains, in the tool that fits your stack (Cube, dbt metrics or LookML)
- MCP server with typed tools, schemas and tests, deployed in your environment
- Agent runtime wired to your chosen interface (chat, Slack, IDE), with explanations and citations back to the metric definitions
- Evaluation suite: golden questions, scoring, a dashboard for accuracy and cost per query
- Auth and policy layer: SSO, row- and column-level rules, audit log
- Monitoring and alerts for latency, error rate and spend
- Runbooks, architecture notes and a handover session with your team
How an engagement runs
Week 0
Readiness scan: which questions people actually ask, which tables answer them, what already has definitions.
Weeks 1–2
Semantic model for the first domain; MCP server skeleton; first 30 golden questions. Stop-point: you can end here and keep the model.
Weeks 3–5
Agent runtime, access control, evaluation loop, second domain. Internal users on it daily.
Weeks 6–8
Hardening: cost ceilings, monitoring, runbooks, handover. Production for a defined user group.
After the pilot: stop, extend to more domains, or move to a dedicated pod. See how we work.
Questions we get
Do we need to migrate our warehouse first?
No. The design runs over the Snowflake, BigQuery or Databricks you already have. If the data model is not ready for a semantic layer, the readiness scan says so before anyone pays for a pilot.
Which models and interfaces do you support?
The MCP server is model-agnostic. It works with the Claude and OpenAI APIs, and the agent can be reached from Claude, ChatGPT, Slack or a code editor. You choose; the governance below the agent does not change.
How do you keep the agent from inventing numbers?
It cannot write free-form SQL. Every answer is a query against the semantic layer, which maps to a defined metric. The evaluation suite catches regressions, and the answer cites the definition it used.
What about people who should not see certain rows?
Access control is enforced in the auth and policy layer and in the warehouse, not in the prompt. The agent runs as the user, with the user's permissions.
What does the pilot leave us with if we stop?
The semantic model, the MCP server, the evaluation suite and the docs, in your repositories. All of it is usable without us.
Is there a public demo?
Not a public one. We walk through the reference design, with a live demo, on request. See the reference design.
Ask us about your warehouse
Bring the three questions your team asks most. We will tell you what it takes to answer them with an agent you can trust.