Skip to content
Discuss a project
  1. Kiz8
  2. Work
  3. Data context for an AI analyst

Data context for an AI analyst

Kiz8 built a read-only MCP data context layer that helped a pilot AI analyst answer 64 of 69 golden SQL questions correctly (92.8%), with MCP p95 32 ms.

Problem

Undocumented Iceberg catalogs prevented people and LLM agents from querying data reliably.

Approach

Track each claim’s source, evidence and review status in Git. Check claims against live data, compile them into a read-only MCP server, and evaluate comprehension, adversarial traps and golden SQL.

Measured outcome

Pilot: 50/50 comprehension, 24/24 adversarial, 64/69 golden SQL (92.8%); MCP p95 32 ms

Tech stack

Lake/Storage
Apache IcebergD1
Transform
BunTypeScript
Serve
Cloudflare WorkersMCP
Start a conversation

From raw data to better decisions.

Send one paragraph about the data, the decision it should support, or the hypothesis you need tested. We reply with what we would look at first.

Discuss a projectExplore our research ›