Infrastructure · Healthcare
In-country LLM infrastructure for healthcare
In healthcare the safest gain from AI is administrative rather than clinical: correspondence, report formatting, document search. Zero-Shot builds infrastructure that does this work without patient data leaving the institution. Diagnosis, treatment and clinical decision support are out of scope.
What we build
- 01An inference layer on internal hardware that makes no outbound calls
- 02An assistant layer for administrative documents and correspondence
- 03Search across internal procedures, instructions and regulation
- 04A de-identification layer applied before processing, with access logging
- 05Role-based access, quotas and retention rules
How it runs
01
Write the scope boundary first
What the system will not do is written first: clinical decisions, diagnostic suggestions and treatment guidance are excluded, and the boundary is enforced architecturally.
02
De-identification
Identifiers are stripped before processing, and the accuracy of that step is measured with its own test set.
03
Pilot on administrative work
The first deployment targets one process with heavy administrative load, and the time saved is measured.
04
Handover
The institution's IT team operates the system, with documentation and training as part of delivery.
What we measure
- De-identification accuracy, with a target of zero leaks
- Time saved on administrative work
- Access log completeness
- p95 latency and concurrent user capacity
- Usage intensity and user satisfaction
When this is the wrong fit
Diagnosis, treatment recommendations, dosing, triage and clinical decision support are out of scope — we do not build them. We also decline systems that speak directly to patients or answer medical questions. Our work here reduces administrative load and speeds up access to internal knowledge.
Common questions
Do you build clinical decision support?
No. Diagnosis, treatment, dosing and triage are out of scope. These areas fall under medical device regulation and doing them properly requires clinical validation. Our systems work on administrative tasks and document access.
How is patient data protected?
Two layers: the system runs inside the institution and makes no outbound calls, and identifiers are stripped before processing. The accuracy of stripping is measured separately and leaks are reported. Access logging is complete.
Where does it actually help?
The clearest gains are in repetitive administrative correspondence, report formatting and internal procedure search. Those tasks are high-volume and low-risk, which makes them the right starting point. Moving closer to clinical work increases risk, not return.
Other services
What we've written on this
- Why AI agent projects failAgents that impress in a demo stall in production. Four observations on tool design, error handling, handing off to a human, and the cost ceiling.
- Is it safe to send company data to AIWhat happens when company data reaches a language model: provider terms, training use, data protection law, cross-border transfer and self-hosting.
- Making LLM costs predictableWhy AI spend surprises you at month end, and how to control it: model tiering, prompt caching, output length, and per-request ceilings.
Last updated: 2026-09-05
Let's talk about this one
Tell us what you want to build and within two days we'll tell you who takes it and how we'd measure it.