Your team wants to ask questions of your data in plain language, through Claude, ChatGPT or an agent of your own. The first demo looks great. Then the answers turn out to be wrong in ways nobody notices at first: a column misread, a join that double-counts, a date filter that quietly uses information that wasn't available at the time. And nobody is comfortable giving an AI model access to production data.
I make your data ready for AI agents: a semantic layer that explains what the data means, tested query patterns for the tricky business logic, a secure access layer built on the Model Context Protocol (MCP), and a test set that measures how often the agent gets the answer right. Fixed scope, fixed price, and everything runs in your own infrastructure.
An accuracy test set with real questions and expected answers checks every change
Every table, column and metric gets one clear name and one description, maintained with the data models (for example in dbt). The agent sees what a column means, not just its type, and the same definitions serve your analysts and dashboards.
The questions that are easy to get wrong (point-in-time logic, deduplication, the right join between two business dates) get a tested SQL pattern the agent can load when it needs it. Patterns are versioned and tested like code.
An MCP server between the agent and your data: read-only SQL, every query validated before it runs, limits on cost and rate, authentication, and query logging. Error messages are written so the model can correct its own mistakes.
A set of real questions from your team with expected answers. It runs on every change, so you know how good the agent is today and notice when something breaks. Valid SQL is not the same as a correct answer.
We look at your data, the questions your team wants answered, and the risks. You get:
I build the four building blocks with your team, in your infrastructure. You get:
New questions turn into new patterns and test cases, and the agent is re-tested whenever models or your data change. Monthly, cancel any time.
A public MCP server that makes 60+ years of US stock market data queryable in plain English from Claude, listed in the official MCP Registry. It runs on exactly this approach: a semantic layer on dbt, PostgreSQL and DuckDB, a library of tested SQL patterns, read-only SQL with validation before every query, and rate limiting with an open-source module I wrote.
What I learned along the way: What 46,000 production queries taught us about LLM-generated SQL (HackerNoon). Rate limiting module: caddy-jsonrpc-matcher on GitHub.
Agent-ready data is only as good as the data underneath. I've built data warehouses, data lakes and pipelines since 2011 for clients in automotive, chemicals, media, market research and telecommunications, and LLM-based workflows that extracted data from ~120,000 documents for a supply chain platform. See my projects for details.
The MCP server and the data stay in your cloud or data center. Which model you connect, and under which data-processing terms, is your decision; the access layer works the same way with any MCP-capable client.
Anything you can query with SQL. I've worked most with PostgreSQL, DuckDB, Athena and Spark-based data lakes on AWS and Google Cloud.
The assessment will tell you. If the foundation needs work first, I can build that too: data platforms are what I've built since 2011. See services.
Want to know whether your data is ready for AI agents? Book a free 30-minute call. We talk through your data, your use cases and whether the assessment makes sense for you, with no obligation. You can also call me at +49 30 4193 6978 or email hello@crichter.io.
Want to forward this to your team? Here is a two-page PDF summary.