Censor private data & shrink docs to fit any AI prompt.

LoyalCloak is a zero-cloud, privacy-first utility that censors sensitive PII (Personally Identifiable Information), slashes token size, and formats documents before pasting into AI assistants like ChatGPT, Claude, DeepSeek, Gemini, and NotebookLM.

Why LoyalCloak?
Pasting raw documents into cloud LLMs risks exposing private names, emails, phones, cards, SSNs, IPs, or keys. LoyalCloak censors sensitive data and collapses bloated formatting 100% on-device, so any document fits your AI context window.

Why Shrinking Token Size Matters:
• Prevent AI Hallucinations: Repeating headers and whitespace noise degrade reasoning. Stripping fluff focuses the model on core context—fewer hallucinations, better adherence.
• Bypass Upload Limits: Never get blocked by "Prompt is too long". Layout normalization compresses documents by up to 50% so massive files fit in a single prompt.
• Cut API Costs & Latency: Fewer input tokens mean lower API bills for OpenAI, Anthropic, and Google—and faster responses.

Key Features:

100% On-Device Privacy & Zero Telemetry
• Native Rust core processes everything locally on-device.
• Zero network calls: no servers, no tracking, no cloud storage.
• Transient RAM execution: documents and ML inferences are wiped immediately.

NotebookLM Export & Smart Splitter
• Bypass 500k-Word Limits: NotebookLM caps sources at 500k words, and large PDFs often fail to upload. LoyalCloak cleans and prepares docs so they upload reliably.
• Calibrated Word Budgets: Automatically split oversized docs into structured Markdown with clean section breaks (presets for 500k, 100k, 50k, or 25k words).
• Batch Merge: Combine multiple research notes into single, model-ready NotebookLM sources in 1 tap.

AI Deep Scan — On-Device Person-Name NER (Pro)
• Embedded local ML transformer (DistilBERT int8) running via ONNX Runtime.
• Detects and redacts full and partial person names in unstructured text without cloud APIs.
• Reversible Pseudonymization: Replace names with realistic pseudonyms ("Jane Doe" -> "Alex Abbott") and restore originals with 1-tap after the AI responds.

Automated PII Censorship & Custom Rules
• Standard Redaction: Emails, phones, FDA NDC drug codes, and ICD-10 diagnosis codes.
• Pro Deep PII Censorship: Credit cards (Luhn-verified), SSNs, IBAN accounts, IPv4/IPv6, API keys, secrets, national IDs, and custom regex rules.
• Pro Pattern Library: Checksum-validated health & legal IDs (NPI, DEA, MRNs, court dockets, account numbers) with per-detector toggles.
• FRCP 5.2 Masking (Pro): Keep only the last 4 digits of SSNs, cards, and account numbers for legal filings.
• Allowlist: Keep trusted terms or corporate domains untouched.

Privacy Audit Report (Pro)
• Export timestamped, on-device summaries of redacted items as PDF or Markdown (Summary, Masked, or Full detail).

Fit AI Context Budgets & Token Analytics
• On-device BPE token counting for OpenAI (cl100k/o200k), Claude, and Universal estimation.
• Smart Auto-Chunking (Pro): Split long docs on section boundaries with labels.
• Custom XML Wrappers (Pro): Encapsulate text in structured prompt tags.

Complete Format Support & OCR
• PDF, Word (DOCX), EPUB, HTML, CSV/XLSX, Plain Text, and images via on-device OCR.
• Android share intent: send documents from Files, Google Drive, or Chrome.
• Save & Share: Copy cleaned markdown, save to .md, export for NotebookLM, or share.

Legal & Privacy:
• Privacy Policy: https://doc.loyallogic.org/privacy
• Terms of Service: https://doc.loyallogic.org/terms