01
Deterministic scrubber
Regex and dictionaries for NIF/DNI, IBAN, case numbers, court references and other patterns you can explain to a regulator.
Alpha · Coming soon
Nothing leaves uncovered.
Route any prompt to any model — OpenAI, Anthropic, Mistral, Gemini or your stack — with a local pipeline that strips, masks and re-hydrates personal data. The upstream model sees pseudonyms; your firm sees the real names.
OpenAI-compatible gateway · API and UI
What leaves your perimeter
████████ ████ advised ██████████ on the █████ merger.
What the model sees
Person_12 and Org_4 advised Client_A on the Northern deal.
What your team sees back
Maria García and Cuatrecasas advised Acme Corp on the Nordic merger.
Built for regulated work
GDPR-aligned design
EU AI Act readiness
Schrems II & transfer risk
Attorney–client privilege posture
Five stages, one perimeter
Deterministic scrubbing first, then models, then an adversarial judge. Answers come back through the same vault so names and entities reappear only for you.
Local execution · Tunable strictness · Audit trail
01
Deterministic scrubber
02
Discriminative NER
03
Reversible vault
04
SLM rewriter
05
Adversarial judge
06
Your LLM provider
01
Deterministic scrubber
02
Discriminative NER
03
Reversible vault
04
SLM rewriter
05
Adversarial judge
06
Your LLM provider
Re-hydrate response· Deterministic per-session tokens: the same person stays Person_12 across turns. The vault re-hydrates the model reply before anyone reads it.
Inside the pipeline
Each stage has a job. Together they keep quasi-identifiers and edge cases from slipping through.
02
Discriminative NER
Fine-tuned encoder models spot people, organizations, locations and legal entities where rules alone miss context.
03
Reversible pseudonym vault
Deterministic per-session tokens: the same person stays Person_12 across turns. The vault re-hydrates the model reply before anyone reads it.
04
Generative SLM rewriter
A small model tuned with RL handles indirect identifiers and phrasing that still points to a single individual.
05
Adversarial judge SLM
A second small model attacks the masked text for re-identification risk and scores k-anonymity / l-diversity style leakage before anything is sent upstream.
Policy as code
Tune strictness per matter, client or API key — not by retraining. Legal sets the policy; engineering ships the binary.
matter_id: M-2025-014
strictness: high
entities: [PERSON, ORG, CASE_ID, IBAN]
allow_foreign_models: true
judge_min_score: 0.92
Example policy surface
Route to the models you already use
Same gateway, multiple backends. Swap providers without changing your anonymization story.
OpenAI
Anthropic
Google Gemini
Mistral
Cohere
AWS Bedrock
Azure OpenAI
Ollama
vLLM
Where Veil earns its place
From litigation to public-sector disclosure — any workflow where a name in the wrong place is the whole problem.
Litigation drafting
Draft and research with LLMs without client names or opponent details crossing your perimeter.
- Pseudonyms for parties, counsel and courts in every upstream prompt
- Consistent mapping across a long thread so the model does not get confused
- Re-hydration only inside your review surface
GDPR DSR / SAR responses
Generate summaries and extracts for data subjects without leaking other individuals in the same file set.
- Entity-level masking tuned for multi-party records
- Judge pass catches indirect references before export
- Policy variants for HR, health and finance record mixes
M&A due diligence
Let deal teams use frontier models on counterparty and employee data under a controlled pseudonym layer.
- Vault keeps code names stable across thousands of pages
- Swap from Claude to GPT without redoing redaction logic
- Audit log of what left the perimeter and when
Clinical notes (HIPAA-minded)
Summarize and structure clinical text for operations while PHI stays tokenized for external models.
- Strong entity coverage for providers, patients and facilities
- Local pipeline keeps raw notes off third-party training paths
- Strictness presets for research vs. operations
Finance / KYC analysis
Run LLM-assisted review on transactions and customer files with account tokens instead of live identifiers.
- IBAN, tax IDs and internal customer handles in the scrubber layer
- Per-desk policies for trading vs. compliance
- Re-hydration gated behind your existing access control
Public sector / FOIA
Prepare releases and internal briefings with consistent redaction before any cloud model sees the text.
- Citizen and official identifiers masked to stable tokens
- Schrems-minded routing: choose EU-only backends per policy
- Documentation-friendly: explain scrubber rules in plain language
Join the waitlist
Alpha access, implementation partners and design reviews. Tell us who you are and what you want to run through Veil.
Indicative — final at GA
How pricing will look
Open core first. Hosted and enterprise when you need someone else to run the judge models and SLAs.
Community / OSS
Free
Self-host the gateway, models and policies. Full source for the pipeline and UI.
- ·Open-source core under a permissive license
- ·Run on your hardware or VPC
- ·Community support and docs
Hosted
From €299 / month
Managed multi-tenant gateway with upgrades, monitoring and predictable scale.
- ·Minerva-operated control plane
- ·Included minor version updates
- ·Standard support windows
Enterprise
Custom
Private cloud, BYOK, on-prem judge and rewriter SLMs, custom policies and contractual terms.
- ·Dedicated or VPC deployment
- ·SLAs and security review pack
- ·Professional services for legal and compliance alignment
Be first in line
We are onboarding a small set of firms and platform teams for the alpha.