AITALON AIMask
AITalon on Telegram

AIMask A safe de-identified version of a document

Local, verifiable, designed to keep source data inside, and natively integrated with AIGate.

MODULE_01 ACTIVE

Local processing

OCR, NER, visual analysis, and masking run inside the protected perimeter without external AI APIs.

ON-PREMOCRNER
SAFE PIPELINE
MODULE_02 ACTIVE

Independent verification

A separate verifier re-extracts content and scans for residual data; a candidate is never treated as a safe result.

RE-SCANDLPSAFE
SAFE PIPELINE
MODULE_03 ACTIVE

Exact, auditable delivery

Profile, hash, and artifact version stay pinned at every stage, while download rechecks access and verdict.

SHA-256RBACAUDIT
SAFE PIPELINE
A SIMPLE FILE JOURNEY

How AIMask cleans a file

A file enters through the interface, API, or MCP, is recognized and cleaned locally, and comes back as a ready anonymized or masked copy. AIGate can use this route natively.

RESULT The source stays inside the protected environment while the clean copy returns through the interface, API, or MCP. FILE CLEAN / READY
AIMask architecture

Five separately deployable services isolate the control plane, file stream, untrusted parsing, local ML, and independent verification. AIGate invokes this perimeter natively before the model call and receives only the published clean version.

WEB / API / AIGATE
mask-gatestream + exact grants
mask-workersandboxed transform
mask-verifierindependent re-scan
admin-maskprofiles · state · audit
ocr-masklocal OCR · layout · logos
A candidate can cross the delivery boundary only after a strict SAFE verdict.
AM

admin-mask

CONTROL

UI, identity, profiles, job state, outbox, retention and content-free audit. The only PostgreSQL schema and domain-write owner.

MG

mask-gate

DATA

Streaming ingress and download authorization with exact-version artifact grants; no database credentials.

MW

mask-worker

SANDBOX

Sandboxed format-aware parsing, detection, masking, sanitization, and immutable candidate creation.

OCR

ocr-mask

ML

Local OCR, layout, and logo inference with immutable models and no document-storage credentials.

VR

mask-verifier

VERIFY

Independent re-extraction, OCR, residual scan, and optional DLP proposal before SAFE publication.

PG

Metadata, not documents

PostgreSQL is authoritative for state. Binary files stay in encrypted artifact storage, never in PostgreSQL or Valkey.

ID

Exact immutable binding

Every stage is bound to tenant, job, generation, profile revision, artifact version, and SHA-256.

Safe-result conditions
LOCAL
Models and OCR
STRICT
Profile and version
SAFE
Only after verification
SB

Parser isolation

CONTROL

The parser child receives no network, database, Valkey, object-storage, or KMS credentials.

HX

Hostile document ingress

HIGH

Signature, MIME, archive limits, traversal, XML/XXE, active SVG, macros, and external links are validated.

VR

Independent verification

CRITICAL

The verifier does not accept worker findings as proof and repeats extraction, OCR, and residual scanning.

DP

DLP fail-closed

CRITICAL

Required ICAP DLP errors cannot silently promote a candidate to SAFE.

LG

Content-free telemetry

CONTROL

Raw document text, OCR output, PII, object keys, and delivery tokens are excluded from logs and metrics.

DL

Revalidated delivery

HIGH

Download checks current identity, object scope, SAFE state, generation, exact artifact, hash, expiry, and one-time token.

Managed policies
PF

Versioned profiles

Detectors, actions, DLP, verification, and retention are pinned in a signed revision for the whole job.

RB

SSO and capability RBAC

OIDC, LDAP/LDAPS, group mapping, and explicit capabilities for users, admins, and security viewers.

RT

Retention and deletion

Policy controls retention; delivery always revalidates current state and exact-version availability.

AU

Explainable audit

Transitions, decisions, manual corrections, and content access are audited without storing source values.

AIMask control plane
Isolated services 5
Core roles 3
Profile revisions per job 1
Raw values in logs 0
AIMask capabilities

One managed pipeline for text, tables, scans, and images, strictly within the certified format registry.

Document de-identification

#01

Detect and irreversibly remove PII, secrets, names, logos, and matches from custom rules.

PIIRULES

Hidden-content sanitization

#02

Remove metadata, comments, revisions, external links, macros, and other hidden or active content according to policy.

SANITIZEPROFILE

Scans and visual regions

#03

Local OCR and layout models find text regions and logos, while visual masking changes the actual output pixels.

OCRVISION

Policy-driven human review

#04

A user can add a mask or justify a false positive. Every change creates a new generation and repeats verification.

REVIEWAUDIT

Native AIGate integration

#05

AIGate uses the built-in file route to send a file to AIMask before the LLM and proceeds only with an independently verified clean version; failures are fail-closed.

AIGATENATIVE

Secure result delivery

#06

A short-lived one-time token binds tenant, job, generation, exact version, and hash of the published SAFE artifact.

ONE-TIMESAFE
ORDER PILOT→ BLBlog
FAQ

AIMask FAQ

Does AIMask send documents to external AI services?

No. The target architecture runs OCR, entity and logo detection, masking, and repeat verification locally, without internet access or external AI APIs.

When is a result considered safe?

Only after an independent mask-verifier pass and required DLP inspection when enabled by the profile. A warning or partially processed file never receives SAFE status.

Which formats are supported?

AIMask uses a format registry and exposes support only after full conformance testing for the exact format and mode. Unknown, damaged, encrypted, or uncertified variants are rejected with a clear reason.

Can AIMask be integrated with AIGate?

Yes. In file-based LLM flows, AIGate sends the document to AIMask before the model call and receives only a verified result. Production routing must fail closed on errors or timeouts.

Where can AIMask be deployed?

Target options include on-premises, private cloud, and a dedicated managed installation. Air-gapped delivery is designed to avoid runtime model or package downloads. Each production topology still requires acceptance of its PostgreSQL, Valkey, storage, mTLS/KMS, and ingress controls.