AccAnalysisAccAnalysis
Revenue Cycle & Financial Operations

Ask the question; no SQL, and no data leaving the network

Non-technical staff could not query account data without writing SQL or waiting on the data team. We built a natural-language interface on a locally hosted model — because the data could not leave the building.

At a glance
Status
Delivered
Region
United States
Engagement model
Dedicated team

Status: Delivered. Described from our own delivery records. Client identity, brand and commercial terms are withheld. Figures, where given, cover the period stated and nothing beyond it.

The organisation

Context

A revenue-cycle operation holding account-level data covering healthcare patients.

Operational and commercial staff regularly needed answers from that data. They could not write SQL, and the data team's queue was the only route.

The brief

The problem

  • Every ad-hoc question became a ticket, and the answer arrived after the moment that prompted it.
  • The data team's capacity was consumed by one-off queries.
  • Self-service BI covers the anticipated questions; it does not cover the unanticipated ones, which are often the important ones.
  • The data includes personally identifiable and health-related information, so any approach that sent it to a hosted model was out of scope before it was discussed.
The work

What we built

A locally hosted large language model

A thirty-billion-parameter model deployed on the client's own on-premises server, with nothing traversing the network boundary.

Schema grounding

So the model translates a question into a query against the actual database structure rather than a guessed one.

A prompt interface accepting voice or text

Returning results in seconds.

Method

How we delivered it

  1. 1

    Establish the boundary first

    What may run where, confirmed with the client's compliance position before any model was selected.

  2. 2

    Provision on-premises infrastructure

    Sized for the model.

  3. 3

    Ground on the schema

    Iterating against a set of real questions staff had previously asked the data team.

  4. 4

    Constrain what the interface may do

    Read paths only, scoped to what the asking user is permitted to see.

  5. 5

    Pilot with a small group

    Of non-technical users, measured against the answers the data team would have given.

Sequence

How it was phased

PhaseDurationWhat happens
1Boundary & model selection
1–2 wks

Residency constraints, model choice, hardware sizing

2Infrastructure
2 wks

On-premise deployment, serving, monitoring

3Schema grounding
3–4 wks

Structure mapping, query generation, evaluation set

4Interface & permissions
2–3 wks

Voice and text input, read-only scoping, user permissions

5Pilot
2–3 wks

Non-technical users, answers compared against the data team's

Indicative phasing for work of this shape. Actual duration varies with data quality, access and decision speed.

Hand-over

What the client keeps

  • The deployment and its configuration
  • Schema grounding artefacts
  • The evaluation question set
  • Permission scoping rules
  • Infrastructure documentation
Stack
On-premise LLM (30B)SQLSchema groundingVoice inputPrivate server deployment
Keep reading

Related

Related case studies

Tell us what your current system can't do, and we'll tell you what it would take to change that.