Self-hosted AI analyst writes SQL, self-checks numbers, and cites every claim to its query

Outside-Risk-8912 · reddit · 2026-08-18

A developer released a self-hosted, source-available (Elastic License 2.0, not OSI open source) AI Analyst built on the principle "show the SQL or it didn't happen" — the opposite of black-box "chat with your data" tools.

Workflow in five steps:

States its approach before touching data: reads the schema, plans steps, and explains why — e.g. explicitly noting the semantic model lacked a hires metric so it fell back to the raw monthly table
Runs each step as readable SQL with row counts and a "where these numbers came from" breakdown
Self-checks every result: in a live HR-dataset test it flagged a genuine 4-person inconsistency between summed net adds (+17) and headcount delta (+13), carrying the caveat into the write-up
Cites findings: every claim references its step, e.g. "headcount climbed 122 to a 140 peak (step 1, step 2)"; verdict: 55% of Engineering hires were net growth, Support is shrinking (backfill ratio 1.42)
Closes the loop: mark verified / flag buttons, suggested follow-ups, scheduling, CSV/PDF export

Stack: one Docker command + your own Supabase; BYOK with any provider (demo ran Kimi K3 via OpenRouter) — the author argues the plan → SQL → check → cite structure, not frontier models, does the heavy lifting. It's part of a larger self-hosted platform (agentswarms) with multi-agent swarms, RAG, BI dashboards, budgets, and tracing.

Original post →

Self-hosted AI analyst writes SQL, self-checks numbers, and cites every claim to its query

More from coding & agent