HIPAA-conscious clinical AI

Clinical AI.
Without the PHI risk.

Tokenize patient data before it ever touches a language model. The LLM sees placeholders like [NAME_001], never real PHI. Responses are re-hydrated locally — so clinicians get the output they need, without data leaving your infrastructure.

AWS BAA-covered
Zero PHI to the LLM
Model-agnostic
clinitect.ai/app
Your transcript

Patient John Doe, DOB 05/14/1968, MRN 12345.

Reports headache x3 days, seen by Dr. Chen.

Tokenized →
What the LLM sees

Patient [NAME_001], DOB [DATE_001], MRN [ID_001].

Reports headache x3 days, seen by [NAME_002].

Re-hydrated ←
Generated Note

John Doe, a 58 y/o male seen on May 14, presents with a three-day history of headache. Evaluated by Dr. Chen

Token map held in memory only · destroyed after response

Built on infrastructure trusted by every major health system

Amazon Bedrock
AWS Comprehend Medical
AWS Textract
AWS BAA

Architecture

Protect, tokenize, reason, re-hydrate.

Every request flows through the same three-stage pipeline — verifiable, auditable, and built on AWS services covered under the same BAA your hospital already trusts.

PHI detected & tokenized

AWS Comprehend Medical + Clinitect regex catches names, dates, MRNs, SSNs, addresses, NPIs, insurance IDs. Every entity gets a deterministic token. Nothing is guessed.

Model never sees patient data

The LLM receives only tokens. Real values stay in your AWS account, in memory, for the lifetime of a single request. No training, no logging, no third parties.

Response rehydrated locally

Tokens in the model output are swapped back with real values before the note ever leaves your server. Clinicians get a complete, accurate note — not a redacted one.

How it works

Secure PHI in. Intelligent output out.

Six steps. Every request. No exceptions.

1

Paste PHI from EHR

Provider pastes protected health information into the extension or web app.

2

Scrape & tokenize PHI

The engine identifies, scrubs, and tokenizes every PHI entity on-device.

3

Send tokens to LLM

Only tokenized content is sent to the model (Bedrock / Claude / Llama / Nova).

4

LLM generates response

The model processes the tokenized prompt and returns structured output.

5

Re-tokenize & re-hydrate

Tokens are mapped back to the original PHI securely, on the server.

6

Deliver output with PHI

The clinician receives a complete response with PHI restored in place.

Zero retention

PHI never leaves your infrastructure

Token maps live in memory for the lifetime of a single request, then are destroyed. Nothing is written to disk. Nothing is logged. Nothing is sent beyond the AWS region you deploy into.

Use cases

Built for clinical workflows that can't compromise on privacy.

VA TBI evaluations

Turn an hour-long veteran interview transcript into a VA-ready, structurally-correct TBI writeup in under a minute.

45 min → 60 sec

Clinical note generation

Convert dictated or typed encounter transcripts into Chief Complaint / HPI / Assessment / Plan structure, ready for the EHR.

~90% adoption-ready

Document OCR + structuring

Upload a scanned intake form or a PDF. Clinitect runs OCR, de-identifies, and converts the content into a structured note.

PDF → note in 3 clicks

Security & compliance

HIPAA you can trust, because you can verify it.

Every architectural choice is visible, logged, and auditable — so your compliance team has answers the first time they ask.

See it in action

End-to-end privacy

PHI is never sent to an LLM in raw form — not Claude, not GPT, not anyone. Tokenization happens before the first network hop.

On-device protection

Tokenization and re-hydration happen inside your AWS account. The token map never leaves the machine handling the request.

LLM agnostic

Works with Amazon Nova, Claude, Llama, and more. Swap models with a single config change — no architectural rework.

HIPAA-conscious design

Runs on AWS services covered by the standard BAA. No third-party vendors in the request path. Full audit trail optional.

Start tokenizing PHI in under 60 seconds.

No signup. No setup. Paste a transcript, pick a model, watch the tokens flow. All without sending a byte of PHI to an LLM.