Product

AI and advanced analytics

Forecasting, machine learning and natural-language querying, built on data that is already correct. These projects succeed or stall on the data layer rather than the model, so that is where we start.

What we take on

Where the value usually is

Dashboards live under Development. This page is about the work that goes beyond reporting what already happened.

AI readiness assessment

An honest look at whether your data can support the use case you have in mind, and what it would take if it cannot. Sometimes the answer is not yet.

Grounded assistants

Natural-language querying over governed data and your own documents, returning the query or the source alongside the answer so it can be checked.

Document extraction

Invoices, contracts and forms turned into structured records, with confidence thresholds and a human review path for anything uncertain.

Forecasting

Demand, cost and usage models built on reconciled history, with assumptions written down where the business can argue with them.

Anomaly detection

Catching the transaction, cost line or data load that does not look like the others, and routing it to someone who can act.

Evaluation and guardrails

Test sets, accuracy thresholds and monitoring agreed before launch. If we cannot measure whether it works, we do not ship it.

The honest part

Natural language is earned, not bolted on

Asking a question in plain English and getting a reliable answer depends almost entirely on the modelling underneath. Certified tables, agreed metric definitions and clear column descriptions are what make generated queries correct rather than merely plausible. We treat those as part of the build, not documentation to write afterwards.

  • Every answer returns the query or source document behind it
  • Scope limited to certified tables - nothing else is visible to it
  • The same permissions as the rest of the platform, not a separate model
  • Accuracy measured against a real test set before anyone relies on it

Fit

When this is right, and when it is not

A good fit when

  • Your reporting layer is already governed and reconciled
  • There is a specific decision the model would change
  • Someone owns the outcome and can judge whether it is working
  • You need a defensible answer on feasibility before committing budget

Probably not when

  • The underlying data is not yet governed - start there
  • The real need is a dashboard, which is Development
  • Nobody can say what a good answer would look like
  • The goal is to have AI rather than to solve something

Tools

PythonSQLRetrieval and embeddingsVector searchEvaluation harnessesForecastingAnomaly detectionIn-database MLModel tracking

Have an idea worth testing?

Tell us the use case and we will say plainly what it takes to make it real, and whether the data is ready for it.

Start a conversation