· NETBRAND CORP
Velô scope and limitations
What Velô detects and what it misses, with what margin of error, at what speed and under which conditions. Figures from our internal evaluation on real clinical documents.
Where the figures come from
From an internal NETBRAND evaluation on real clinical documents in Spanish that were not used to train the model. Velô's data, methodology and technical details are a trade secret and are not published.
Results and margin of error
| Metric | Measured | Range we assume | What it means |
|---|---|---|---|
| Coverage: text with personal data that gets redacted | 97.1% | 95% to 99% | Out of every 100 characters of personal data, 1 to 5 may remain visible |
| Patient data coverage: name, ID number and date of birth | 98.9% | 97% to 99.9% | The best-protected data, but not 100% |
| Documents with any character of a patient data point visible | 9.1% | 6.6% to 11.6% | About 1 in 10 documents needs review if zero leaks is required |
| Documents with a serious leak: half or more of the data point visible | 8.7% | 6.3% to 11.1% | Most leaks are fragments, not the full data point |
| Precision: redacted text that really was personal data | 90.4% | 88% to 92% | About 1 in 10 redacted characters was not personal data |
| Clinical content redacted by mistake | 0.25% | up to 0.5% | Medical terms are almost always preserved |
How to read the margin. The ranges add an extra margin to the evaluation's statistical margin, because every institution writes and structures its documents differently and text quality varies by format.
Leaks are small but spread out. Only 1.8% of patient data stays partly visible, but those fragments appear in 8.7% of documents.
Results by data type
F1 combines two things: that the data point is detected and that it gets the right type. A low F1 does not always mean the data stays visible: it may be redacted under another type or with a different boundary.
| Data type | F1 | Reliability |
|---|---|---|
| Phone | 0.988 | High |
| Web and IP | 0.988 | High |
| Age | 0.978 | High |
| Patient name | 0.963 | High |
| Patient ID number | 0.949 | High |
| Date of birth | 0.949 | High |
| Clinician name | 0.936 | High |
| 0.932 | High | |
| Address | 0.891 | Medium |
| Clinician license number | 0.867 | Medium |
| Institution | 0.846 | Medium |
| Record identifiers | 0.831 | Medium |
| Other people: relatives, companions | 0.824 | Medium |
| Location | 0.564 | Low |
Location is the hardest type. Velô redacts 97% of its characters, but often under another type or with imprecise boundaries. For medium- and low-reliability types we recommend sampled human review.
Training data through 2026 and new entities
The model was trained on data available through 2026. Anything that appears later, or that was not represented in training, may not be identified well. For example:
- New identifier formats: a new numbering for national IDs, medical records, health plan membership or clinician licenses.
- New names of institutions, insurers, programs, services or places.
- Uncommon names, nicknames, abbreviations or spellings.
If you find cases like these, report them: they can be covered with rules in your installation or in a later version of the model. We recommend measuring again whenever your document sources change.
Languages and markets
Velô recognizes personal data and identifiers from Colombia, the United States and Brazil. Performance on your documents is measured in the POC, before going to production.
Text only: not OCR
Velô processes text, not images or audio. The figures were measured on the text of the documents. If a PDF contains scanned pages or images, Velô detects it and quarantines it, unprocessed, so it can be reviewed another way. Personal data inside an image is not detected. Tables, headers and forms, where data appears without context, are also harder.
Processing speed
- 10,000 to 50,000 pages per hour per device with NVIDIA GB10, such as the Velô Nodo or the NVIDIA DGX Spark. This is the range we recommend for capacity planning.
- What moves the number: text density (more words per page, fewer pages per hour) and actual GPU availability (if the device is shared with other workloads, such as a local language model, speed drops). File conversion and read/write speed also matter.
What Velô does not promise
- It does not detect 100% of personal data. Risk goes down; it does not disappear.
- It does not process images and is not OCR. PDFs with images are quarantined.
- The figures on this page are internal evaluation results, not a guarantee for each document. Performance on your documents is measured in the POC.
- Tokenizing with a reversible vault is pseudonymization, not anonymization. The data is still subject to personal data regulation.
- It is not a certification for HIPAA, GDPR, LGPD or Colombia's Law 1581 of 2012, and it does not replace the assessment of your privacy committee or legal counsel.
How to operate within this scope
- Run a POC on your own hand-annotated documents and measure coverage per data type before going to production.
- Review by sampling, weighting location, other people and record identifiers, and decide who handles quarantined documents.
- Minimize: send the AI model only the fields it needs.
- Report missed cases so they make it into the next version.
Aggregate results from an internal NETBRAND evaluation, November 2026. This page contains no personal data and is not legal advice.