Can we use it with any model?
Yes. Nakato is model-agnostic.
It works with any LLM API, so you keep the model you have chosen. Whichever model you use, it only ever sees your data's twin.
- Models
- Any LLM API
- Sits
- Inline, inside your environment
Your AI never meets your customer. It meets their twin.
Nakato is a privacy layer that swaps sensitive data for contextually faithful twins. A model can reason with realistic data without you sacrificing privacy.
Synthetic example. No real customer data appears on this page.
See the swap in four sectorsHow it works
Semantic
Contextual detection
Nakato reads the whole document and finds every sensitive detail, direct or indirect. This includes indirectly identifying information that pattern matching misses.
Pseudonymisation
Twin generation
Every sensitive value is swapped for a semantic twin: fictional, but contextually faithful. Same shape, different person.
Reversible
Perfect reversal
When the answer comes back, the twin is reversed back to the original data. Every substitution is logged and auditable. Real data has never left your environment.
| Type | Original | State |
|---|---|---|
| person | Margaret Ellis | found |
| date of birth | 1961-04-12 | found |
| account number | 20-45-11 88451203 | found |
| address | 14 Rowan Grove, Sheffield | found |
4 values found, read for meaning rather than matched.
RSP is our data transformation pipeline. It converts sensitive details into realistic twins at the moment of query, so the model gets the context it needs to work. Nakato runs inside your VPC or on-premises. Your data stays put; the only thing that ever leaves your environment is the twin.
The demo
Real decisions. Fictional identities.
One layer, wherever regulated data is the blocker. Pick a sector: the record on the left stays in your environment, the model only sees the record on the right.
Banking
Credit decisions with context
Record · Banking Twinned at query time
| Field | Original stays with you | Twin what the model sees | Preserved |
|---|---|---|---|
| Customer | Margaret Ellis | Eleanor Grant | Preserved — gender, name origin |
| Date of birth | 1961-04-12 | 1959-11-03 | Preserved — age band |
| Sort code | 20-45-11 | 31-08-64 | Preserved — valid format |
| Balance | £18,240.55 | £17,905.12 | Preserved — within 2% |
Credit decisions with context. Stripping demographics strips the vulnerability signals credit analysis depends on most. The significance of each detail is preserved, meaning your AI tool can reason properly. No employee or client is identifiable.
The model sees the right-hand column but the answer you read has the originals restored. Where the two columns match is where details essential to the task are kept: RSP keeps what is essential to the task's reasoning but makes it unidentifiable.
Evidence
Our numbers.
The field is the benchmark. Each point stands for the same number of test cases. Each mark reaches as far as that tool's figure.
- Placeholder masking
- Nakato
- The rest
The field is the benchmark. Each point stands for the same number of test cases.
- Nakato
- The rest
The field is the benchmark. Each point stands for the same number of test cases.
- Nakato
The field is the benchmark. Each point stands for the same number of test cases. Each mark reaches as far as that tool's figure.
- Nakato
- GLiNER
- The rest
Summarisation accuracy on the same cases. A longer bar is better.
- Nakato
- Placeholder masking 41.3%
- Tokenisation 12.8%
Leak rate under the same attack. A shorter bar is better.
- Nakato
- GLiNER 7.8%
- Microsoft Presidio 22%
Shown alone. A comparison is drawn only where the same benchmark produced one.
Checked across 500,000 financial-services test cases. The full report, with its methodology, is available on request.
How each figure was measuredIn line with the rules you already work to.
Nakato is built to the regulator's own test for pseudonymised data, so your compliance team can see how it meets the rules before anyone asks.
Where does our data go?
Your real data never leaves your environment.
The model never receives a real value, and the mapping remains in your infrastructure. Nothing about your existing data classification has to change.
- Runs in
- Your VPC or on-premises
- Model
- A self-hosted small language model
Does it meet GDPR?
Designed to the regulator's own test.
The model receives only a twin it has no reasonable means of tracing back to a person, which is the test GDPR applies when deciding whether data identifies someone.
- The test
- Means reasonably likely to be used (Recital 26)
- The boundary
- Your pseudonymisation domain
Can we prove it?
Every substitution is auditable.
Every swap is written down - what was replaced, with what, and when - so a reviewer can retrace a decision without ever seeing the underlying data.
- Trail records
- Each transformation and its reversal
- Ready for
- Supervisory review
Built against the EU AI Act, DORA, FCA and PRA expectations, GDPR and SOC 2 from the start.
What the team building it will ask.
Security signs off on where the data goes. Whoever runs the AI programme wants to know whether it still works. These are the questions we hear most often.
If the model sees fiction, how is the answer still right?
The twin keeps what the reasoning needs.
A twin changes what identifies a person but keeps what matters to the task: an age band, a format, a balance within a few per cent. The model reasons with equivalent context, and the answer comes back to the user with the originals restored.
- Summarisation accuracy
- 96.2%
- Reversal
- 100%, every substitution
What does it add to latency?
Tens of milliseconds.
Detection and twinning happen at the moment of each query. There is no batch job to wait for. An audit log is kept inside your environment.
- Added latency
- 45–90ms
- Runs
- At query time
Another question? Write to hello@nakato.ai and we will answer it.
Join us
Work on the frontier.
Generating semantic twins is not a solved problem. It takes an unusual mix of linguistics, semantic ontology and computer science to achieve it. If you’re attracted to the idea of working with us to keep developing this solution, talk to us.
Semantic ontology meets computer science. Two disciplines, one twin that keeps its meaning.
The problem changes shape with every domain. What preservation means for medical data looks nothing like insurance.
The work does not repeat. Small team, hard problems, and no template to follow.
We intend to define how this is done.