Why RedactiPHI?

We offer enterprise-grade de-identification accuracy (95% F1 score) at a developer-friendly price point, plus features that others charge extra for or don't offer at all:

• Joinable Tokenization - Same patient = same token across all documents. Enable longitudinal analytics without exposing PHI.
• One-Click Re-identification - De-identify for LLM processing, then restore original values. Critical for AI scribes.
• Cryptographic Audit Receipts - Tamper-proof proof of what was processed, when, and what was found.
• Developer Dashboard - Usage analytics, API key management, logs - not just a CLI or library.
• Built-in Webhooks - Get notified when async processing completes.
• 5-Minute Setup - pip install, get API key, start de-identifying. No Spark clusters or on-prem servers.

Private AI is excellent - they support 50+ entity types, 52 languages, and have strong accuracy. They're a great choice for enterprises with complex requirements.

We differentiate on:

Choose Private AI if: You need 50+ languages, have dedicated DevOps for container deployment, or have enterprise procurement budget.

Choose RedactiPHI if: You want to start immediately, need joinable tokens for analytics, want transparent pricing, or need re-identification for LLM workflows.

John Snow Labs has the best accuracy in the industry (96-98% F1). Their Healthcare NLP library is incredibly powerful and customizable.

The tradeoffs:

Choose John Snow Labs if: You need the absolute highest accuracy, have Spark infrastructure, need deep customization, or are processing billions of records.

Choose RedactiPHI if: You want competitive accuracy without managing Spark, need a simple REST API, want built-in re-identification, or have moderate volume (<1M docs/month).

Cloud provider APIs are convenient but have limitations:

• Lower accuracy - AWS scores ~83% F1, Azure ~91% F1 vs our 95%
• No re-identification - Once redacted, the data is gone forever
• No joinable tokens - Can't link the same patient across documents
• Vendor lock-in - Tied to their cloud ecosystem
• Usage-based pricing adds up - $0.01/unit sounds cheap until you're at scale

We're cloud-agnostic, offer re-identification, and provide joinable tokens out of the box.

HIPAA note: AWS/Azure are "HIPAA-eligible" but require you to sign a BAA and configure everything correctly. We're HIPAA-compliant out of the box with BAA available.

Microsoft Presidio is a solid framework, but it's exactly that - a framework you need to build on.

• Accuracy gap - Vanilla Presidio scores 70-75% F1. Can be improved with tuning, but that's your job.
• No healthcare optimization - General PII, not clinical text. Misses things like "Dr. Smith" in medical context.
• You manage everything - Hosting, scaling, monitoring, updates, security.
• No re-identification - Build your own token storage and mapping.
• No HIPAA compliance - You're 100% responsible for making it compliant. No BAA available.

Open source is great for learning or highly custom needs. For production healthcare use, you'll spend more on DevOps than our subscription costs.

When we de-identify "John Smith" in document A, we create a token like [NAM_abc123]. When "John Smith" appears in document B (for the same subject), we create the same token.

This is critical for:

• Longitudinal patient analysis
• Training ML models on de-identified cohorts
• Quality metrics across encounters
• Research datasets that need linkability

Re-identification restores the original values from tokens. The primary use case is LLM workflows:

1. Clinical note comes in: "Patient John Smith, DOB 03/15/1980..."
2. De-identify before LLM: "Patient [NAM_abc123], DOB [DAT_def456]..."
3. Send to GPT/Claude for summarization, coding, etc.
4. Get response with tokens: "[NAM_abc123] presents with..."
5. Re-identify for clinician: "John Smith presents with..."

The LLM never sees real PHI, but the final output is human-readable.

Access control: Re-identification requires the document ID and is audit-logged. You control who can re-identify.

Yes! Common use cases include:

• Analytics pipelines: De-identify clinical data, maintain joinable tokens for cohort analysis
• ML pre-training data: Prepare clinical text for model training without PHI exposure
• Data warehousing: Build de-identified data lakes for research teams
• LLM fine-tuning: Create training datasets from clinical notes safely

Every de-identification produces a signed receipt containing:

• Hash of input document
• Hash of output document
• PHI types detected and counts
• Policy applied
• Timestamp
• Cryptographic signature

This creates tamper-proof evidence for compliance audits. You can prove what was processed, when, and that the output hasn't been modified since.

We detect all 18 HIPAA Safe Harbor identifiers plus clinical extensions:

• Names: Patient, provider, family members
• Dates: DOB, admission, discharge, procedure dates (ages 89+ generalized)
• Identifiers: SSN, MRN, insurance ID, account numbers
• Contact: Phone, fax, email, address, ZIP code
• Digital: IP address, URLs, device identifiers
• Clinical: Facility names, provider credentials

On our internal benchmark of real clinical notes:

• Precision: 95.7% (low false positives)
• Recall: 95.2% (low false negatives)
• F1 Score: 95.4%

We use a multi-engine approach: pattern matching, transformer-based NER, medical terminology filtering, and name detection heuristics. Results are reconciled with confidence scoring.

Note: The demo on our homepage uses only the pattern engine for speed. The full API uses all engines.

Currently supported:

• Text: Plain text, JSON, HTML, Markdown
• Documents: PDF (native + scanned via OCR), password-protected PDFs
• Healthcare: FHIR R4, HL7v2, C-CDA/CCD XML
• Archives: ZIP (processes all files inside)
• Images: PNG, JPG, TIFF (via OCR)

Coming soon:

• Password-protected ZIP and 7z archives
• Full C-CDA structured parsing (preserve XML structure)
• EHI export bundles (eClinicalWorks, Epic MyChart, Cerner)
• FHIR Bulk Data ($export) NDJSON streaming
• EMR-specific format handling (Epic, Cerner, Athena, NextGen)

• Encryption: AES-256-GCM for stored data, TLS 1.3 in transit
• Token storage: Token mappings are encrypted and isolated per customer
• No training on your data: We never use customer data to train models
• SOC 2 Type II: (In progress)
• BAA available: For healthcare customers