Reference: Natural-Language Data Analyst

Architectural notes

Reference architecture for a "talk to your warehouse" analyst agent.
Translates natural language to SQL, runs against a read-replica, returns
formatted results + charts to Slack.

Why these picks:
- Sonnet for SQL generation; Haiku for the routing/clarification turns.
- Postgres MCP for transactional warehouses; Snowflake MCP for analytical.
- Read-only credentials enforced at the MCP server layer + guardrails to
  catch destructive queries before execution.
- E2B for chart rendering (matplotlib in a sandboxed kernel).

Common gotchas:
- Schema context dominates accuracy; invest in a docs layer (column descriptions)
  over just-the-DDL.
- Always show the generated SQL before running; surface costs of expensive queries.