AI data privacy and PII masking for industrial teams.
sanitai is an AI gateway for industrial workflows. It masks personal data and sensitive business identifiers before model calls, enforces governance policy, supports secure RAG, and restores context locally when policy allows it. The public playground demo is live, supports French and English only, and is still in development.
Paste your text. Watch the safe version appear live.
This is no longer a developer console. Visitors can test SanitAI with their own text, run a real sanitization request, and watch the protected output stream in before copying or reviewing it. This public demo is still being refined and currently supports French and English only.
The protected version appears here, token by token, as soon as you start the sanitization flow.
This experience focuses on the transformation itself. The output represents what can cross the trust boundary without exposing raw values.
Four clear dispositions for every request.
sanitai does not treat every model call the same. Each prompt or document is classified into one of four operating modes before it can cross the boundary.
Content with no protected personal, business, or industrial identifiers. It can be sent to an approved provider without placeholder substitution.
- Public product documentation
- Generic support instructions
- Non-identifying business reporting
Content whose sensitivity is carried by identifiers rather than by the procedure itself. Protected values are replaced locally with deterministic placeholders before dispatch.
- Supplier contact records and account ownership
- Purchase approvals with order references
- Part, lot, and supplier mapping documents
Content that may still benefit from AI assistance but should remain inside the company boundary. Use local models or reviewer workflows only.
- Engineering notes with partial process detail
- Assembly review material
- Quality investigations that should remain internal
Content where the secret is the method, formula, tolerance logic, or operational know-how itself. It must not be sent to an external model.
- Manufacturing procedures with exact steps
- Formulas, recipes, and blend ratios
- Calibration settings and critical tolerance tables
Policy is not a slide. It is an execution path.
The same four-tier model drives masking, dispatch, validation, and rehydration decisions across the product.
Discuss your policy modelUseful outputs without uncontrolled exposure.
sanitai stays practical because governance and product utility are designed together. The system protects identifiers without turning AI adoption into a manual process.
Policy decisioning
Every prompt, document, and provider response is evaluated against an explicit policy model before a model call is allowed. sanitai decides whether content can be sent as-is, masked first, kept local, or blocked entirely.
- Deterministic detection for industrial and business identifiers
- Fail-closed handling for ambiguous or high-risk content
- Auditable outbound checks before every provider call
Sanitized retrieval
Documents are sanitized before ingestion so the retrieval layer never indexes raw protected identifiers. Query linking and local rehydration keep answers usable without widening the trust boundary.
- Placeholder registry created before indexing and retrieval
- Queries can still resolve supplier, part, and lot references
- Responses are validated before local rehydration
Operator review lab
Operators can replay traces, inspect provider behavior, and test policies in a controlled environment without turning internal experiments into a production-facing surface.
- Trace replay for provider behavior and policy review
- Operator-owned workflows for experiments and demonstrations
- Separated from production traffic and enterprise request handling
Request early access
We are working with a small set of industrial teams that need AI assistance without letting supplier, quality, or engineering identifiers leak outside their control plane.
Or write directly to hello@sanitai.io
You will talk to the team building the product, not a generic form pipeline.
We onboard a limited number of teams so feedback can influence the roadmap.
Conversations focus on real documents, real policy constraints, and deployment conditions.
Built for teams that cannot treat model calls as ordinary API calls.
Supplier operations, quality teams, engineering reviewers, and governance functions all need the same promise: external models do not see raw protected identifiers unless policy explicitly allows it.
Protect account ownership, supplier contacts, and commercial identifiers while still using external models for summaries and triage.
Review incident narratives, lot references, and escalation records without pushing raw identifiers outside the boundary.
Keep the line between safe operational assistance and blocked process know-how explicit and enforceable.
Separate content that can be masked and routed from content that must remain local because the secret lives in the method itself.