International AI Processing Notice

Lexaris — PII Anonymization Platform
Document Reference: Anon-Legal-INT-29

This document is dynamically generated based on settings configured by the Operator in the Platform Admin Panel. Values shown in italics represent current settings and may be modified by the Operator within legally permitted limits.

---

1. Purpose of This Notice

This notice is issued pursuant to Articles 13(2)(f), 14(2)(g), and 22(3) of Regulation (EU) 2016/679 ("GDPR") and addresses Lexaris (IE Kudryavtsev S.A.)'s ("Controller") use of artificial intelligence ("AI") and automated processing technologies within the Lexaris platform ("Platform").

Lexaris is a PII anonymization platform. Its core function — detecting and anonymizing personal data in documents — relies on multiple layers of automated processing, including AI/ML components. This notice provides meaningful information about:

• Each AI/automated processing layer in the detection pipeline;
• The legal bases under which such processing is conducted;
• The local, open-source nature of all AI components;
• The fact that no data is sent to commercial AI APIs by the Platform;
• Human oversight and the End User's control over anonymization results;
• Data subjects' right to opt out of specific AI layers;
• The Controller's position under the EU Artificial Intelligence Act (Regulation (EU) 2024/1689).

This notice supplements the GDPR Privacy Policy (Document 24) and should be read in conjunction with it.

---

2. AI and Automated Processing Layers

2.1 Overview

The Platform's detection pipeline consists of four processing layers, each serving a distinct function. All layers operate entirely on the Controller's own infrastructure within the Russian Federation. No uploaded document content is transmitted to any external AI service.

2.2 Processing Layers Table

2.3 Detailed Layer Descriptions

Layer 1 — Regex Detection (Deterministic)

The regex layer uses hand-crafted regular expression patterns to detect structured PII:
- Russian passport numbers (series + number with checksum validation)
- SNILS (Social Insurance Number — with control sum verification)
- INN (Taxpayer Identification Number — with check digit validation)
- Phone numbers (Russian and international formats, with normalization)
- Email addresses
- Dates (multiple formats)
- Bank account numbers
- Vehicle registration plates
- IP addresses
- Custom patterns defined by the End User

This layer is fully deterministic: given the same input, it always produces the same output. It does not involve machine learning or AI. Checksum validation reduces false positives for structured identifiers.

Layer 2 — spaCy NER (Statistical ML)

The spaCy layer uses pre-trained Named Entity Recognition models:
- ru_core_news_lg: Russian-language NER model trained on news corpus. Detects PERSON, ORG, LOC, and other entity types.
- en_core_web_trf: English-language NER model based on transformer architecture. Detects similar entity categories in English text.

These models are trained on publicly available datasets and are not fine-tuned on user data. The models run locally; no data leaves the Controller's servers during spaCy processing.

Layer 3 — Ollama LLM (Optional, Explicit Activation)

The Ollama layer provides enhanced NER capabilities using a locally deployed large language model (qwen3:8b). This layer:
- Is disabled by default and requires explicit user activation in account settings
- Provides contextual understanding that statistical NER may miss (e.g., recognizing that "the accused" in a specific context refers to a named individual mentioned elsewhere)
- Runs entirely on the Controller's own GPU hardware within the Russian Federation
- Does not transmit any data to external servers
- Uses an open-source model (qwen3); the Controller does not use proprietary/commercial LLMs
- Does not retain or learn from user data — each request is stateless

Layer 4 — EasyOCR / Tesseract (ML-Based OCR)

For image-based documents (scanned PDFs, photographs of documents, screenshots):
- EasyOCR: PyTorch-based OCR supporting Russian and English text recognition
- Tesseract: Traditional OCR engine as fallback
- OCR output (extracted text) is fed into Layers 1-3 for NER processing
- All OCR processing is performed locally; no images are transmitted externally
- GPU acceleration available on the Controller's servers (for EasyOCR)

---

3. Critical Distinction: No Data Sent to Commercial AI APIs

3.1 Platform Architecture

Unlike many AI-powered services, Lexaris does not transmit any user data to commercial AI providers such as OpenAI, Anthropic, Google, Microsoft, Yandex, or any other external AI service. The entire detection pipeline operates on the Controller's own infrastructure using open-source models.

This architectural decision has significant implications for data protection:
- No international transfer of document content to US-based AI providers
- No risk of user data being used for AI model training by third parties
- No dependency on third-party AI provider privacy policies or terms of service
- Full Controller custody of all data at all times

3.2 End User's Independent Use of Anonymized Output

End Users may independently choose to submit anonymized output documents (i.e., documents with PII replaced by tokens) to external AI services for further processing (e.g., analysis, summarization, translation). Such use:
- Is the End User's own decision and responsibility
- Is outside the scope of the Platform's processing activities
- Benefits from the Platform's anonymization: the submitted text contains tokens ([PERSON_1], [ORG_1]) instead of real PII, significantly reducing privacy risk
- The Platform is specifically designed to enable this use case safely

---

4. Legal Bases for AI Processing Under GDPR

4.1 Layers 1-2 and 4 (Regex, spaCy, OCR) — Core Functionality

(a) Performance of a Contract — Article 6(1)(b):
These layers are integral to the anonymization service. Processing is necessary to perform the contracted service: detecting PII in uploaded documents. An End User cannot receive the anonymization service without these processing layers operating.

4.2 Layer 3 (Ollama LLM) — Optional Enhancement

(a) Consent — Article 6(1)(a):
Ollama processing is optional and requires explicit user activation. The activation flow clearly describes what Ollama processing involves (local LLM analysis of document text for enhanced entity detection) and obtains consent before enabling the feature. Users may disable Ollama processing at any time.

(b) Performance of a Contract — Article 6(1)(b):
Where the End User has activated Ollama processing as part of their use of the Platform, processing is also justified as necessary for contract performance (the user has opted into this as a feature of the service).

4.3 Special Category Data — Article 9

Uploaded documents may incidentally contain special category data (health information, political opinions, religious beliefs, etc.). The Platform processes such data solely for the purpose of detecting and anonymizing it — a protective purpose. Legal bases:

• Article 9(2)(g): Substantial public interest — data protection compliance
• Article 9(2)(a): Explicit consent of the End User who uploads the content

---

5. Human Oversight and End User Control

5.1 End User as Decision-Maker

The Platform's NER pipeline generates detection results and suggested anonymizations. The End User retains full control:

(a) Review Before Finalization:
All NER detection results are presented to the End User for review before anonymization is finalized. The End User can:
- Accept or reject individual entity detections
- Add entities missed by the pipeline
- Modify entity type classifications
- Adjust anonymization tokens
- Choose which entity types to anonymize and which to leave unchanged

(b) No Autonomous Action:
The Platform does not autonomously distribute, file, or act upon anonymized documents. The End User decides what to do with the output.

(c) Confidence Indicators:
Each detection result includes a confidence score. Low-confidence detections are flagged for End User attention.

5.2 Output Labelling

Anonymized documents are clearly labelled as machine-processed. Detection reports indicate which AI layers contributed to each entity detection.

5.3 Professional Responsibility

End Users are responsible for reviewing anonymization results before relying on them. The Platform is a tool; the End User exercises judgement about whether anonymization is adequate for their specific purpose. See the GDPR Privacy Policy (Document 24) Section 14 regarding limitations.

---

6. Opt-Out Granularity for Each AI Layer

6.1 Available Opt-Out Options

6.2 Per-Session Override

End Users can override their default pipeline configuration on a per-session basis. For example, a user who normally has Ollama enabled can disable it for a specific document, or vice versa.

6.3 Right to Object Under GDPR Article 21

Where AI processing is conducted on the basis of legitimate interests (Article 6(1)(f)), data subjects may exercise their right to object under Article 21 GDPR. Objections should be submitted to privacy@lexaris.ru.

---

7. Data Protection Impact Assessment

7.1 DPIA for AI Processing

The Controller has conducted a Data Protection Impact Assessment (DPIA) pursuant to Article 35 GDPR for the Platform's AI processing activities, specifically:

• Automated PII detection in uploaded documents using ML models (spaCy NER);
• Optional LLM-based processing (Ollama);
• OCR processing of image documents;
• Processing of documents potentially containing special category data.

DPIA outcomes are documented and available to the competent supervisory authority upon request.

7.2 Ongoing Assessment

DPIAs are reviewed:
- Before deploying a new NER model or updating existing models;
- When the scope or scale of AI processing materially changes;
- Following a data breach or security incident;
- At least every two years.

---

8. EU Artificial Intelligence Act Readiness

8.1 Overview

Regulation (EU) 2024/1689 (the "EU AI Act") entered into force on 1 August 2024, with provisions applying progressively through 2026.

8.2 Risk Classification Assessment

The Controller has assessed the Platform's AI systems against the EU AI Act risk classification framework:

(a) Low-Risk Classification:
The Controller's position is that Lexaris constitutes a low-risk AI system under the EU AI Act. The Platform:
- Is not listed in Annex III (high-risk AI systems)
- Does not assist in the administration of justice (Annex III, point 8)
- Does not perform biometric identification (Annex III, point 1)
- Does not make decisions affecting natural persons' access to essential services
- Is a privacy-enhancing tool that assists users in protecting personal data, not in making decisions about individuals

Anonymization tools are not categorized as high-risk AI systems under the EU AI Act. The Platform's purpose is protective: it helps users comply with data protection law by removing PII from documents.

(b) General Purpose AI (GPAI) Models:
The Platform incorporates a GPAI model (qwen3:8b via Ollama) deployed locally. Where applicable, the Controller complies with transparency obligations under Article 50 EU AI Act. The model is open-source with published documentation.

(c) Prohibited Practices — Article 5:
The Controller confirms that it does not deploy AI systems falling within the prohibited categories: subliminal manipulation, exploitation of vulnerabilities, social scoring, or real-time remote biometric identification.

8.3 Transparency Obligations

In accordance with EU AI Act transparency requirements:
- All NER detection results are labelled as machine-generated
- Detection confidence scores are visible to End Users
- This notice provides meaningful information about each AI layer
- Model provenance and training data categories are documented
- End Users are informed that outputs require human review

8.4 Comparison with High-Risk AI Systems

To clarify Lexaris's position, the following comparison illustrates the distinction:

---

9. Model Provenance and Training Data

9.1 Open-Source Models

All AI/ML models used by the Platform are open-source with publicly available documentation:

9.2 No Training on User Data

None of the models used by the Platform are trained, fine-tuned, or updated using user-uploaded documents or any user data. Models are used in inference mode only. Model updates are applied from upstream open-source releases after internal testing.

9.3 No Proprietary/Commercial Models

The Controller does not use proprietary or commercial AI models (GPT-4, Claude, Gemini, YandexGPT, or similar). All AI components are open-source, auditable, and run locally.

---

10. Updates to This Notice

This notice will be updated to reflect:

• New AI capabilities or NER models added to the Platform;
• Changes to the detection pipeline architecture;
• Published guidance from the European Data Protection Board or European Artificial Intelligence Board;
• Changes to the EU AI Act's applicability timeline or scope;
• Data Protection Impact Assessment outcomes.

Material updates will be communicated to End Users with at least 30 days' notice. The current version of this notice is always available at lexaris.ru/legal/ai-processing.

---

11. Contact

For questions about AI processing activities:

• Email: privacy@lexaris.ru
• Online: lexaris.ru/legal/ai-processing
• Postal: Data Protection Officer, Saint-Petersburg, Russia

---

Configurable parameters

The following parameters are configured by the Operator in the Platform Admin Panel. Values shown below are defaults.

All parameters are configured in: Admin Panel → Settings → Retention Periods

---

This notice constitutes a draft requiring review and approval by qualified legal counsel, the DPO, and technical personnel responsible for AI systems before publication. EU AI Act classification assessments should be reviewed against Commission guidance and EAIB opinions as they become available.

Document Reference: Anon-Legal-INT-29 | Version Date: 2026-03-19