Nema Veil
Nema Veil is our own model for detecting personal data. It finds names, companies, places and contact details in text – so Nema Crew can replace them with placeholders before anything goes to an AI model.
- Version
- 3.0.0
- Based on
- GLiNER2 (mDeBERTa-v3)
- Runs
- on Nema’s servers in the EU
What it detects
Veil marks spans in a text and says what they are: a person, a company, a place, an email address, a phone number, an IBAN, an ID or customer number, a date of birth or health information. Crew does the replacing and translating back. Brands, acronyms, field names and code identifiers should stay readable – “Google”, “SEO” or send_mail are not personal data.
Where it runs
Veil runs on Nema’s own servers in the EU. No text goes to another provider for detection. Together with the open base model GLiNER2 it also detects text in images when that is switched on in Crew.
Benchmark
Measured on our own test set: 375 texts in German and English – chats, agent replies, emails, tickets, tool results and code. They contain 1,010 real details that must be detected and 1,456 traps that must stay readable. No text of the test set appears in the training data; the benchmark checks this itself.
| Metric | Veil 2.0 | Veil 3.0 |
|---|---|---|
| Leaksreal detail not detected · lower is better | 6 (0.6%) | 5 (0.5%) |
| Trap hitsbrand, acronym or field name replaced · lower is better | 375 (23.8%) | 90 (5.7%) |
| False detectionsper 1,000 characters · lower is better | 6.6 | 1.7 |
| Over-maskingshare of replaced characters without a detail · lower is better | 9.0% | 2.9% |
| Exact spansstart and end match · higher is better | 89.2% | 90.2% |
| Compute timeper 1,000 characters, CPU · lower is better | ≈ 930 ms | ≈ 920 ms |
On a second test set with 1,358 details, Veil 3.0 detects 98.2% (Veil 2.0: 98.3%) and over-masks 4.5% (before: 5.7%).
As of 23 October 2026
How we measure
- Our own test set
- We wrote the test set ourselves, modelled on the kinds of text that occur in Crew. It is not public and not an independent test.
- A fixed release rule
- A new version only goes live if it has no more leaks than the old one, overall and in every category, hits traps at most half as often, and has no more false detections.
- No word lists
- Veil detects from context, not from a fixed list of names or brands. Each organisation maintains its own terms in Crew.
Where it has limits
- Veil doesn’t catch everything: in the benchmark, about one in 200 details slips through. Whatever it misses goes out unchanged.
- We measured German and English. We have no numbers for other languages.
- We tuned some thresholds on the second test set, so it is no longer entirely untouched.
- Veil still takes some well-known brands for customers’ companies, such as “Amazon Web Services”.