DecisionBox
Autonomous AI Discovery For Your Data — Open Source
Start new thread
Selçuk Kızıltuğ

DecisionBox Enterprise - Agent that writes SQL for you to validate database insights

by
DecisionBox
•
DecisionBox is an autonomous AI agent that writes SQL against your warehouse and ships validated findings. Enterprise runs it fully air-gapped: self-hosted LLMs via Ollama, open-source base models fine-tuned on your schema, SSO, RBAC, three-layer data governance, full audit log. Plugin architecture on an open-source AGPL v3 core — zero fork, zero outbound calls, zero bytes leave your network.

Add a comment

Replies

Best
Selçuk Kızıltuğ
Maker
📌
Hey everyone, Selcuk here, one of the founders of DecisionBox. We open-sourced DecisionBox a few months ago. It's an AI agent that connects to your data warehouse, writes its own SQL, and finds things you didn't know to ask about. The open-source version is solid and people are using it. But every conversation with a company that wanted to try it went the same way: "looks great, how does auth work? Can we control what the agent sees? Do you have audit logs?" So we built the Enterprise edition. Same discovery engine, but wrapped with the stuff orgs actually need. SSO with any OIDC provider, role-based access, data governance (schema filtering, column redaction, query rewriting), full audit logging, and a Slack integration so findings land where your team already works. It's all plugin-based on top of the open-source core. No fork, no separate product. Open-source keeps getting better independently. We're running a free 2-week POC: three validated findings on your own data, no commitment. decisionbox.io/contact Happy to answer anything in the comments; schema, governance, deployment, whatever's useful.
Dhruba Patra

I love the fact this is open source and it gives this product much much more credibility !

Also would like to know how you are verifying claims using LLMs - its a big research topic now a days !

Can Abacigil
Maker

@dhruba_patra Thanks, really appreciate that. Open source was non-negotiable for us. Nobody should have to trust a black box that's touching their warehouse.

On validation, it's not one trick, it's a few layers stacked together.

First, the agent isn't allowed to just report numbers it saw in an exploration query. For every claim it wants to make ("X% of users did Y"), it has to write a separate independent SQL query that verifies that specific number, run it, and check the result matches. If the numbers don't line up, the claim gets adjusted with the verified number or rejected entirely. Both outcomes are visible in the UI so you can see what the LLM originally said vs what your warehouse actually returned.

Second, every insight has to cite the exploration step numbers it came from, so you can trace any finding back to the raw query and results that produced it. No floating claims.

Third, affected counts are forced to be COUNT(DISTINCT user_id) in the verification SQL, which kills a whole class of LLM math errors.

It's not foolproof - an LLM can still pick a misleading angle on correct numbers - but it closes the hallucination gap hard. The honest framing is: the LLM proposes, the warehouse disposes.

Can Abacigil
Maker

Adding one more thing, happy to go deeper on any of these in the thread:

• Air-gapped deployment: Ollama model choices, GPU requirements, outbound-call audit: https://decisionbox.io/enterprise/audit/
• Governance: the three-layer protection (schema filtering, result redaction, AI query rewriting) : https://decisionbox.io/enterprise/governance/

Drop your questions below, happy to answer.