Keep patient data out of the LLM. Prove every clinical decision your AI assisted.
A single piece of PHI in a third-party model prompt is a HIPAA exposure and a dead hospital deal. Swiftward detects and removes patient data before any prompt leaves your environment, and records every AI-assisted decision so you can defend it later.
PHI never leaves your environment
Swiftward runs on your own infrastructure and sits in the path between your application and the model. It detects and redacts patient identifiers and clinical free text before a prompt is sent: deterministic matching for the known identifiers, a model for the free text, and it fails closed. If it is not sure, it blocks the prompt rather than sending it best-effort. When redaction is verified, the PHI never reaches the third-party model, which materially reduces your disclosure surface. On-prem is not a preference here; it is the architecture that makes the rest possible.
The hard part is the combinations
The obvious identifiers are the easy part. The hard part is what re-identifies a patient with no single field to match on: a ZIP, a birth date, and a rare condition mentioned together in a free-text note. That is what the free-text model is for, and why it fails closed. And we will not quote you a redaction rate measured on someone else's data: in the pilot we measure it on your own notes, your team reviews the misses, and you tune it until it clears your bar.
When the answer comes back wrong
Redaction handles what goes in. The other half is what comes back: a recommendation that is confidently wrong. Swiftward does not judge whether the medicine is right, and no honest tool claims to. What it does is make sure a questionable answer cannot pass silently. You write the check (low confidence, no citation to the record, an answer outside the agreed scope), and when it trips, the output goes to a clinician before anyone acts on it, with the whole exchange on the record.
Prove the decision
When an AI assists a clinical decision and someone asks, months later, what it recommended and on what basis, you replay the exact decision on the exact policy version that was live. The record holds the inputs the AI saw, its recommendation, the policy version that was active, any clinician override, and the timestamps. The same record answers an auditor, a regulator, and a malpractice review.
Built for the hospital security review
Your buyers' security teams ask for on-prem, access control, an audit trail, and a Business Associate Agreement before anything else. The foundation is built in: on-prem, SSO, role- and attribute-based access, and append-only audit. And the BAA story is simple: because Swiftward runs on your infrastructure with no data-plane sub-processors, PHI never reaches us. We sign a BAA regardless.
Become a design partner
This is the same engine applied to PHI and clinical-decision audit, and we are taking a small number of healthcare design partners now, with the founder in the room. You would be early, and you would shape how the product handles clinical AI rather than inherit someone else's choices.