Governed AI on your data
Text-to-SQL, assistants and agents over the same governed tables, allowed to see only what the column classification permits, with the model chosen by data-residency policy and the evidence of every call kept in your own tables.
Where you are
Agents and copilots are arriving faster than the controls. Nobody can say which data an answer used, whether it held personal data, or why the model was allowed to see it. Boards want AI; risk wants an audit trail; the vendor's content filter is neither.
How the engagement runs
Governed tables first
Lineage and personal-data classification are produced by the pipelines, so the AI layer inherits them instead of guessing.
Gate every call
Each model call is checked against the column classification and the applicable control before it runs; the decision, the data touched and the model used are written to your own tables.
Model by policy
Bedrock, Anthropic, OpenAI or a local model, selected per task by residency and sensitivity rules. Switching provider is a policy change.
Controls mapped to frameworks
The control library maps to APRA CPS 220 and 234, CPG 235, ISO 27001:2022 and NIST CSF 2.0, so the evidence lines up with what the auditor asks for.
Already built
- ●Text-to-SQL on Databricks and Snowflake, gated by classification, with every call logged.
- ●AI governance control library mapped to APRA, ISO 27001 and NIST CSF; model routing across Bedrock, Anthropic and OpenAI.
- ●LLM observability with OpenLIT on ClickHouse inside the same estate.
What you bring
- ●The frameworks you report against
- ●Your preferred model providers and residency rules
- ●The governed tables, or let the pipelines service produce them