Sleep-therapy provider
≈ $60M revenue (approx.) · 40 clinics · one clinical reporting app
Natural-language clinical reports
Clinicians ask for patient cohorts in plain English and a constrained pipeline turns the question into validated filters, never free-form database queries, on a private GPU that costs a fraction of the hosted model it replaced.
- Organization
- Sleep-therapy provider with a clinical reporting application
- Question
- Could a small open model replace the hosted model in a constrained parser?
- Runs on
- A single cloud project with an inexpensive GPU
- Scope
- Natural-language report queries, validated before execution
Where they started.
Clinicians wanted to ask for cohorts in plain English: patients within a range of adherence, a number of days, a device condition. The existing parser used a hosted frontier model, which meant every question left the provider's environment and cost money per call. The team needed to know whether a private model could match it before committing.
What was built.
- A constrained-intent pipeline: the model emits a fixed schema that is checked against an allowlist of fields before anything runs; never free-form database queries.
- A small open-weight model served on an inexpensive GPU inside the same cloud project as the application, so inference never crosses the boundary.
- A provider seam that lets the application switch between the hosted model and the private one with an environment variable.
- The provider's own benchmark questions replayed against both to settle the question with evidence.
What changed.
- The open model reproduced the hosted model's answers on nine of ten benchmark questions; the tenth was unverifiable in the original logs.
- Every query valid against the schema; no invented fields.
- Inference at a fraction of the hosted model's cost, and the GPU can be paused when idle.
- A dress rehearsal for larger private deployments on the same pattern.
Clinical data belongs inside the provider's boundary. Keeping the parser's model in the same project made the privacy claim mechanical rather than contractual.
A single cloud project, private inference endpoint
Components involved.
More like this.
Multi-brand industrial manufacturer
≈ $200M revenue across five brands (approx.) · 5 brands · 4 ERPs · 1 CRM
- Use Case:
- Consolidation, forecast and pricing intelligence across four ERPs
- Results:
- One governed lakehouse behind an AI agent: audit-grade consolidated financials that trace to the source transaction, forecast and pipeline by business unit, large-job margin watch, and plain-English inquiry over all of it, inside the company's own Azure tenant.
Healthcare receivables agency
≈ $10M revenue (approx.) · ~60 employees · national client base
- Use Case:
- Placement-file intake automation
- Results:
- Placement files arriving in any layout are normalized into the agency's existing upload format, with a human approval queue before anything reaches the system of record. A one-and-a-half-person manual job becomes a review step.
Healthcare receivables agency
≈ $10M revenue (approx.) · ~60 employees · regulated healthcare receivables
- Use Case:
- Protected-health-information send gate
- Results:
- Every outbound file that carries protected health information is checked against the placement it belongs to and held for human approval, with an audit trail of who released what, and when.
Have one like it?
Every spotlight started as a conversation about a process nobody liked doing, under a data-locality constraint.