De-identification for the age of AI.
DeIdentify & Label PHI.
DelPHI detects and redacts protected health information across clinical text, DOCX and PDF — while a clinical-code shield preserves the ICD-10, CPT and HCC codes your pipeline depends on. From offline CPU to GPU-accelerated cloud.
DelPHI is an assistive de-identification control — pair it with a human QA checkpoint for Safe-Harbor sign-off.
Pt: Maria Gonzalez, 58F. MRN 004821. DOB 03/14/1967.
Seen 04/02/2026 for poorly-controlled T2DM E11.9 and CKD stage 3 N18.30. A1c 9.2%. Started metformin.
Tel (415) 555-0147. PCP Dr. Alan Reyes.
Illustration with synthetic data. The web app demo uses synthetic data only — do not upload real PHI.
DeIdentify & Label PHI
DE.L.PHI
DelPHI is a clinical de-identification engine from Jivica AI. It reads clinical text, DOCX and PDF, finds protected health information in context, and removes it — while a clinical-code shield keeps the ICD-10, CPT, HCC and NDC codes your downstream pipeline depends on.
It runs in three tiers: Lite (offline on CPU, the entry tier), Full (offline on GPU, highest accuracy — coming soon), and Web (managed on GCP, highest accuracy, BAA, with FHIR/DICOM connectivity coming soon). DelPHI is an assistive control: pair it with a human QA checkpoint for Safe-Harbor sign-off.
Data is the fuel for AI. PHI can't flow freely.
Every health team wants clinical text in their LLMs, analytics and coding pipelines. Raw notes are full of identifiers — so the data can't safely or legally move. De-identification is the control that lets you say yes to AI without exposing patients.
The AI shift cuts both ways
AI finally makes unstructured notes usable at scale — and simultaneously makes re-identification easier, stitching quasi-identifiers back together. That raises the bar for the de-identification step, not lowers it.
Why generic tools fall short
- Free text breaks regex — keyword tools over-redact or miss identifiers.
- Naïve redaction destroys the codes, dates and intervals analytics need.
- Much clinical data is scanned image-PDF that needs OCR.
- Air-gapped sites, residency law or no BAA appetite block cloud-only tools.
Secondary use of PHI needs Safe Harbor (18 identifiers) or Expert Determination.
→ DelPHI targets all 18 Safe Harbor categories — de-identified data falls outside PHI rules.
Consent, purpose limitation and data-fiduciary duties on personal & health data — with data-residency pressure.
→ On-device de-id (Lite) keeps data in-country and shrinks the regulated footprint.
Pseudonymisation is a recognised safeguard; data minimisation by default.
→ DelPHI's consistent pseudonymisation maps directly to it — intervals and joins survive.
Three clear stages
The same detect → shield → redact workflow powers all three tiers — Lite, Full and Web.
Detect
A clinical NER engine reads context — not keywords — to find all 18 HIPAA Safe Harbor identifier categories across free text: names, MRNs, dates, contacts, geographies and more.
Shieldthe IP
The DelPHI clinical-code shield preserves ICD-10, CPT, HCC and NDC codes plus medical terminology — so the output stays safe and usable in front of a coding or analytics pipeline.
Redact
Apply your mode: Anonymise to type labels, Pseudonymise to consistent surrogates (dates shift by one offset so intervals survive), or Delete identifiers entirely.
One workflow, shared by every edition
Lite, Full and Web run the same de-identification feature set — detect → shield → redact. What scales across tiers is compute and accuracy (CPU → GPU), plus Web's cloud data connectivity.
18 HIPAA categories
Every Safe Harbor identifier category in the detection surface — not a regex subset.
Clinical-code shield
Preserves ICD-10 / CPT / HCC / NDC codes and medical terminology while redacting true PHI.
Three redaction modes
Anonymise to labels, pseudonymise to consistent surrogates, or delete identifiers entirely.
Presets + denylist
PHI-type presets plus a custom term denylist for organisation-specific identifiers.
Documents + OCR
TXT, DOCX and PDF — including scanned image-PDFs handled via OCR.
Side-by-side review
Compare original vs de-identified, with PHI removed broken down by category.
Audit-ready export
Export to TXT · DOC · PDF · JSON, including a structured report for your audit trail.
Intervals survive
Consistent pseudonymisation shifts dates by a single offset so time-to-event survives.
Who reaches for DelPHI
Feed AI safely
Clean clinical text before LLM training, fine-tuning, RAG or eval sets.
HCC & analytics
De-identify in front of coding and analytics — the code shield keeps ICD-10/CPT/HCC intact.
Research & trials
Share datasets for research and clinical trials without exposing identities.
Outsourced coding & QA
Let offshore or external teams work on de-identified documents.
Health-tech SaaS
Embed de-identification inside your own product via the same engine.
Ad-hoc desk work
A clinician or analyst de-identifies a file before it leaves the building.
One engine, three ways to run it
All three share the same detect → shield → redact engine, all 18 categories and three modes. They differ on two axes — where it runs (your device vs our GCP cloud) and how much compute (CPU vs GPU) — and Web adds enterprise data connectivity an offline desktop can't.
DelPHI Lite
Offline on your CPU — the fastest, lightest way to start.
- Runs
- On your device · CPU
- Accuracy
- High (compact clinical model)
- Privacy
- 100% offline · no BAA needed
- Install
- Any Windows PC, no GPU
- Data
- TXT · DOCX · PDF
- Best for
- Getting started · small practices · air-gapped
DelPHI Full
The larger model on your GPU — top accuracy, still 100% offline.
- Runs
- On your device · GPU (NVIDIA / CUDA)
- Accuracy
- Highest (larger clinical model)
- Privacy
- 100% offline · no BAA needed
- Install
- Desktop + CUDA libraries
- Data
- TXT · DOCX · PDF
- Best for
- Air-gapped orgs that need top accuracy
DelPHI Web
Highest accuracy in the browser — plus FHIR/DICOM connectivity.
- Runs
- GCP cloud · GPU (managed)
- Accuracy
- Highest (larger clinical model)
- Privacy
- Cloud · BAA-covered for production
- Install
- None — any device, in the browser
- Data
- Files + FHIR & DICOM (coming soon)
- Best for
- Health systems · EHR/PACS · scale
Demo uses synthetic data only — do not upload real PHI.
De-identification that speaks FHIR & DICOM
Web composes with Google Cloud's Healthcare API and Sensitive Data Protection to de-identify structured records and medical images — while DelPHI's clinical-grade engine handles the free-text notes generic tools get wrong, preserving ICD-10 / CPT / HCC / NDC codes instead of shredding them. Compose, don't compete.
- FHIR records
- DICOM imaging
- Google Cloud Healthcare API + DLP
- Code-safe free-text notes
What the desktop tiers can't match is the cloud connectivity — not the accuracy. Full matches Web's accuracy offline; Web adds the integrations.
Start with a pilot, not a price sheet
DelPHI is in beta — Lite is free to seed adoption, Web is volume-priced at GA. The fastest way in is a focused pilot on your own documents: Lite for the can't-egress case, a Web sandbox for cloud teams. We agree success criteria up front, with a human-QA checkpoint, and credit 100% of the pilot toward a Year-1 subscription.
Say yes to AI — without exposing patients
Join the beta. Tell us where you'll run DelPHI and we'll set you up with Web access, the Lite download, or a pilot on your own documents.