NL-to-SQL analytics that business teams actually trust.
We built a natural-language analytics system over 500+ warehouse tables that reaches 90%+ accuracy and is used daily by 40+ business users. The hard part is not generating SQL; it is knowing when the answer is wrong. Our system is designed around that.
What it is
A multi-agent architecture splits the job: one agent resolves what the question means against a semantic layer, another plans the query, another writes and validates SQL, and a checker verifies the result against known metrics before anything reaches the user.
The semantic layer is built from your schema, metric definitions and past queries, so 'active users last quarter' means the same thing to the model as it does to finance. Ambiguity triggers a clarifying question, not a guess.
Every answer carries the SQL, the tables used and a confidence signal, with full audit logs. Dashboards, scheduled questions and Slack delivery sit on top.
Who it's for
- Data teams buried in ad-hoc requests from business users
- Operations, marketing and finance teams who need numbers without waiting for a ticket
- Companies on Databricks, Snowflake, BigQuery or Postgres with a large schema
- Leadership that wants self-serve analytics with controls
What you get
- Semantic layer with metric definitions and table documentation
- Multi-agent query system with validation and confidence reporting
- Chat, dashboard and Slack interfaces with role-based access
- Evaluation set from real questions, with measured accuracy
- Audit logs, cost controls and a handoff your data team owns
How it works
- 01
Schema
We map tables, metrics and the questions people actually ask, and build the first semantic layer.
- 02
Pilot
A pilot with one team and a measured accuracy target on their real questions.
- 03
Harden
Validation agents, clarifying questions, guardrails and access controls.
- 04
Roll out
More teams, scheduled questions, dashboards and ongoing accuracy tracking.
Related work
Multi-agent analytics
Natural language to SQL over 500+ tables for a leading consumer media platform. 90%+ accuracy, used daily by 40+ business users.
In-house ad-ops platform
Replaced a $200K+/yr SaaS tool with a custom platform: maker-checker workflows, real-time budget alerts and auto-pause rules across ad networks.
Questions
- How do you get above 90% accuracy?
- A semantic layer, multiple specialised agents, validation against known metrics, and an eval set of real questions that we test every change against. Accuracy is measured, not claimed.
- Will it expose data people should not see?
- No. Access follows your existing warehouse permissions and row-level rules, and every query is logged.
- Which warehouses do you support?
- Databricks, Snowflake, BigQuery, Redshift and Postgres in production so far.
Related services
AI Agents Development
Multi-agent systems and agentic automation that run in production, with evals and guardrails.
AI Content Intelligence
Models that read whole catalogues, predict hits and write commissioning reports.
Generative AI Consulting
From strategy to a production system, with a team that runs its own AI products.
Let's scope it.
Tell us the outcome you need. You'll get a concrete plan, a fixed quote and a prototype date.
Start a project