Healthcare & life sciences
Fact-checking & guardrails for clinical AI assistants
Healthcare AI fails in two expensive ways: sensitive data goes out, or a wrong number comes back. A patient identifier that reaches a model provider is a reportable disclosure; a dosage the model subtly altered is a safety event. Both failure modes hide inside otherwise-useful answers.
The proxy screens PHI into reversible placeholders before any request leaves for the model – names, identifiers, contact details never reach the provider – and restores the real values only on the way back to your application. What the model sees, and what any log stores, is screened.
On the way back, deterministic numeric grounding recomputes arithmetic and matches every figure in the answer to the source, so an altered dosage or lab value is caught without relying on another model’s judgment. Content-safety classifiers flag unsafe or out-of-scope advice, and logging is tiered: screened by default, raw only as a deliberate, audited opt-in.
Where it breaks
- ✕PHI leakage to the provider
- ✕Hallucinated dosages / botched figures
- ✕Unsafe or out-of-scope advice
What answers it
- → PHI screen & restore (placeholders)
- → Deterministic numeric grounding
- → Content-safety classifiers
- → Screened-by-default logging
The altered dosage
A summarization assistant restates a medication plan and drifts one figure – 15 mg becomes 50 mg. Numeric grounding matches every number in the answer against the source document, flags the mismatch deterministically (no judge model involved), and in prevent mode blocks the response before a clinician ever sees the wrong figure.
Does patient data reach the LLM provider?
No. PHI is detected and replaced with placeholders before any data reaches the model provider. The provider only sees screened text – placeholders, not real PHI. The unscreened values are returned in responses sent back to your application.
How do you catch a hallucinated dosage?
Deterministically. The grounding stage recomputes arithmetic and compares every figure in the answer to the source material. A number that drifted from the source is flagged or blocked – no LLM judgment required for the numeric class of errors.
What gets stored in logs?
Screened transcripts by default – placeholders, not PHI. Raw storage is available as an opt-in for compliance documentation, but all logs are stored in the US. Screened transcripts are the default for data minimization.
Can it block unsafe medical advice?
Content-safety classifiers (e.g. ShieldGemma-class models) score responses and block or withhold unsafe content, alongside topic controls for out-of-scope advice.
See it on your own traffic.
We’re running a limited demo – sign up and we’ll get you in as soon as we can.