Reference architecture · Illustrative

How a multi-tenant SaaS platform gets answers it can trust.

A worked example, not a specific client, of the kind of production system Meroo is built for: a 200+ table database, per-client custom fields, and strict tenant boundaries, where a wrong answer is worse than no answer.

200+
tables in a schema like this
27
custom fields for a single client
Weeks
from kickoff to a live assistant
90%+
accuracy target*
The challenge

Non-technical teams can't self-serve.

Operations and account staff need everyday answers: how many accounts are active, which customers are near a usage limit, what a client's plan includes. But every question means waiting on someone who can write SQL against a sprawling schema.

Off-the-shelf AI fails on two counts. It's inaccurate: asked “how many accounts are currently active,” a generic model queries the wrong table entirely, because nothing tells it that status lives in a specific column with specific values. And it's unsafe: nothing structurally guarantees one customer's data can't appear in another customer's answer.

Regional health system
What curation looks like

Curation, then guardrails.

A relationship map

Real schemas rarely have clean foreign keys. Relationships are implicit. Curation captures how tables actually relate, so the assistant never has to guess a join.

A business glossary

The terms your business uses, mapped to your data, so statuses, archived rows and tenant boundaries mean what your team means, not what a model assumes.

Client-specific custom fields

Each client defines their own fields. Curation makes them legible to the assistant, so a question like 'seats used against plan' returns a real number rather than a guess.

Tenant-safe execution

Read-only access and tenant isolation enforced in code, independent of the model. Every request is validated before it runs. Unsafe ones are refused, not attempted.

The outcome

From guesswork to a trusted assistant.

Staff ask in plain English and get a written answer plus the underlying data, scoped to the right client, every time. Questions that a generic model gets wrong return correct figures, and the generated query is always available for anyone who wants to check the work.

The difference isn't a bigger model. It's that the assistant finally understands the data the way the team does.

(the Meroo thesis, and what curation delivers)

*Illustrative target: the accuracy curation is built to reach on a schema like this, not a measured client result.

See it answer questions from your own data.

A 30-minute call. We'll show a live demo and, if it's a fit, scope a Discovery engagement.