Your sensitive data never leaves the building.

Velô Nodo: a compact graphite-aluminium AI computer with a micro-perforated front, on a stone plinth.

Detect and transform the personal information in medical records, PDFs, images and audio, inside your own network.

Detect, transform, return. All on your premises.

Every medical record is an identity document.

And the model you want to summarise it with lives in the cloud. Velô stands between the two: it finds every piece of personal data and transforms it before the text crosses your firewall.

Patient and companion namesNAME
SSN, driver's license, member IDID
Record and authorization numbersRECORD
Street address and ZIP codeADDRESS
Insurer, hospital and attending physicianFACILITY
Admission, discharge and follow-up datesDATE

Live redaction

Detect
Tokenize
Synthesize
Send

A discharge summary exactly as it leaves the EHR. This is how you would see it in a PDF.

Discharge summary · EHR export to PDFFictional data

Northfield General Hospital · Burlington, VT Patient: Mary Frances Rogers. SSN: 412-58-7390. Age: 67. MRN: 00482137. Insurance: Blue Ridge Health Plan, member ID BRH-5582041. Address: 1452 Maple Ave, Apt 3B, Burlington, VT 05401. Phone: (802) 555-0142. Emergency contact: George Rogers (son), george.rogers@mail.com. Admitted: 03/14/2026. Discharged: 03/19/2026. 67-year-old female with COPD GOLD 3 and hypertension, presenting with five days of progressive dyspnea. Evaluated by Dr. Louise Meyer (NPI 1548392017), who started bronchodilators and systemic corticosteroids. Improved during the stay; SpO2 90% on 2 L/min oxygen at discharge. Prior authorization #88213-4 requested from Blue Ridge Health Plan for home oxygen therapy. Follow-up on 04/02/2026 with Dr. Andrew Vance at the Northfield General Hospital outpatient clinic.

It reads U.S. clinical notes.

NER models trained on clinical notes, rules for the identifiers generic models miss (SSN, MRN, NPI, member IDs), and an optional second pass with an LLM running on the same Node. Every detection is logged with its type, position and confidence for audit.

Colombia · United States · Brazil

Close-up of the Node's micro-perforated front with its status LED lit.

Identity: Given names and surnamesNAME; Social Security number412-58-7390; Driver's license and passportD1234567; Age67 years
Clinical: Medical record numberMRN 00482137; Insurer, member ID, facilityFACILITY; Prior authorization#88213-4; Clinician NPINPI 1548392017
Contact: Street address1452 Maple Ave; City, state and ZIPVT 05401; Mobile and landline(802) 555-0142; EmailEMAIL
Time: Numeric dates03/14/2026; Written-out datesApril 2
Financial and digital: Cards and accountsPAN, IBAN; IP, URL, usernameDIGITAL
Custom: Your own types: internal codes, site names, sequence numbersCUSTOM; Trained from annotated examples in the UI, no data science team requiredUI

Six ways to treat each value. Per entity type.

Pick a policy per entity and per project: names get synthesized, dates get shifted, the SSN gets tokenized.

Mode	What it does	Example
Redact	Replaces the value with its type label.	Mary F. Rogers → [NAME]
Tokenize	A consistent token across the whole document, reversible with the local vault.	Mary F. Rogers → NAME_1
Synthesize	A realistic value of the same type and format; the text still reads naturally.	Mary F. Rogers → Claire I. Morgan
Shift dates	One offset per patient, so clinical intervals are preserved.	03/14 → 04/08
Generalize	Age to a range, ZIP to its first three digits, to lower re-identification risk.	67 years → 65 to 69
Format-preserving encryption	Encryption that keeps the format, reversible only with your key.	412-58-7390 → 689-21-4053

Any file goes in. It comes out clean.

Out in the same format, or as Markdown and JSON ready for RAG, fine-tuning and analytics.

PDF

Input: Native or scanned, with local OCR
Output: A new PDF with no hidden text layer, metadata or attachments

Office

Input: DOCX, PPTX, XLSX
Output: The same file, structure and styles intact

Data

Input: CSV, JSON, XML, HTML, plain text
Output: Free-text fields cleaned; structured columns by policy

Images

Input: PNG, JPG, TIFF, photos of prescriptions and orders
Output: Text found by OCR and covered at the pixel level

Audio

Input: WAV, MP3, contact-centre calls
Output: Local transcription, silence over every sensitive segment

Email

Input: EML and MSG, attachments included
Output: Body and attachments processed in a single pass

In batches from a folder or a bucket, or one document at a time over the API. The same engine either way.

Velô Nodo.

An AI computer the size of a book. Plug it into your network, it boots without internet and processes your documents on site.

Memory: 128 GB unified CPU + GPU memory
Storage: 4 TB NVMe, encrypted at rest
Network: 2 × 10 GbE, air-gapped mode
Throughput: 3,000 to 10,000 pages per hour
Size: 15 × 15 × 5 cm
License: Offline validation, no outbound calls

Reference specifications, subject to change.

Already have servers? Velô Motor installs on your cluster with Docker or Kubernetes, on x86 or ARM, with a 16 GB NVIDIA GPU or CPU only.

It plugs into what you already run.

A REST API and Python SDK for your pipelines, a web UI for reviewers, and connectors that read and write where your documents already live.

Python · illustrative exampleSDK

from velô import Motor

motor = Motor("https://velo.intranet")   # your Node, inside your network
doc = motor.procesar("discharge_00482137.pdf",
                     modo="tokenizar", idioma="en-US")

respuesta = llm.generar(doc.texto)          # the model only sees tokens
print(motor.rehidratar(respuesta, doc.boveda))

Connectors

Amazon S3
Azure Blob
Google Cloud Storage
SharePoint
SMB and NFS shares
SFTP
API REST

Governance

Roles and permissions per project, SSO with SAML or OIDC
An audit log of every detection and every re-hydration
Sampled human review from the UI
Recall per entity type, measured on your own validation set

Regulatory framework

HIPAAU.S.
GDPREuropean Union
LGPDBrazil
Law 1581Colombia
Res. 1995Colombia

Designed to support compliance with these frameworks. These are not certifications or official seals.

Pseudonymization is not anonymization. If a vault can reverse it, what travels is still pseudonymized data. We help you document the control for your committee and pair it with zero retention and a DPA from your cloud provider.

Weeks 1 to 3

On-site POC

With 100 to 200 real PDFs from your most common sources, hand-annotated, we measure recall per entity type.

Custom

Tuned to your EHR

Rules per export format (Epic, Oracle Health, MEDITECH and others) and entities specific to your institution.

Production

Node or cluster

We install the Node or the Motor in your infrastructure, with 8×5 or 24×7 support and offline model updates.

Use any model you like. It will never meet your patient.

Your networkCloud

Indischarge_00482137.pdf · 3 pages

Out to the model[NAME_1], [AGE_1], COPD GOLD 3, five days of dyspnoea. Discharge plan?

ReplyHome oxygen therapy for [NAME_1]; follow-up in two weeks.

Re-hydrated in your networkHome oxygen therapy for Mary Frances Rogers; follow-up in two weeks.

The vault never leaves the Node. The cloud provider receives tokens or substitutes; the answer comes back with your real data only inside your network.

Try it with your own text.

Paste a note. It is processed in your browser with demo rules for U.S. clinical notes; nothing is sent to any server. The product uses NER models, not just rules.

Input

Output

Pricing. Licensed per device, no page limits.

You pay for each device Velô runs on, not per document or per page. Process everything your hardware can handle.

With hardware

Velô Nodo + license

USD 7,000 per device, one-time

Preconfigured and ready to plug into your network.

First month of license free
From the second month: USD 4,500 per month per device

Example, 6 months: USD 29,500
Paying all 6 months upfront: USD 26,125

Device USD 7,000 + 5 months of license. The 15% discount applies to the license.

Your hardware

License only

USD 4,500 per month per device, from the first month

For companies that already own compatible hardware: any system with NVIDIA GB10 Grace Blackwell and 128 GB of unified memory, such as the NVIDIA DGX Spark or GB10 systems from other manufacturers.

Example, 6 months: USD 27,000
Paying all 6 months upfront: USD 22,950

6 months of license with a 15% discount for paying upfront.

No limit on pages or documents. The only limit is the number of devices.
3,000 to 10,000 pages per hour per device on average: fewer when documents need OCR, more when they are native digital.
6-month minimum term on either option.
15% off the license when you pay the full term upfront.
50% off the second license when you buy more than one device.
Priority support included while the license is active.
Shipping outside the U.S. not included. Quoted separately by destination.
Prices in U.S. dollars (USD), per device.

Request a POC

Measure it on your PDFs.

An on-site POC, with your documents, and nothing leaves your network. You finish with recall per entity type and a report for your committee.

HIPAAU.S.
GDPREuropean Union
LGPDBrazil
Law 1581Colombia

Designed to support compliance.

Frequently asked questions

What is Velô?

Velô is PII detection and de-identification software for clinical text that runs on-premise: on the Velô Nodo, a compact AI computer, or on your own cluster. It processes medical records, PDFs, images and audio without the data leaving your network.

Does my data leave my network?

No. Detection, transformation and the reversal vault all run inside your infrastructure. If you choose to use a cloud model, only tokenized or synthesized text travels, and the answer is re-hydrated locally.

Is pseudonymization the same as anonymization?

No. If a vault can reverse the tokens, the data is still pseudonymized and still subject to personal data regulation. Anonymization requires that re-identification is not reasonably possible, and that has to be measured.

Which markets and identifiers does it cover?

The United States (SSN, MRN, member ID, NPI, ZIP), Colombia (cédula, medical record number, EPS insurer, local address formats) and Brazil (CPF, Cartão SUS, prontuário, CRM, CEP), plus names, dates, phone numbers, emails and institutions.

Which formats does it process?

Native and scanned PDF with local OCR, DOCX, PPTX, XLSX, CSV, JSON, XML, HTML, images, audio and email with attachments. Output can be the same format, or Markdown and JSON ready for RAG.

How much does it cost?

Licensing is per device with no page limits: USD 4,500 per month per device, with a 6-month minimum. With the Velô Nodo (USD 7,000 per device, preconfigured) the first month of license is free. Paying the full term upfront gives 15% off the license. If you buy more than one device, the second license is 50% off. Each device processes on average 3,000 to 10,000 pages per hour, depending on whether documents need OCR or are native digital. Priority support is included; shipping outside the U.S. is not.

How do I start?

With a one-to-three-week on-site POC on 100 to 200 real, hand-annotated documents. You finish with recall per entity type and a report for your privacy committee.