Global Format Coverage
Built-in deterministic rule packs for 19 Countries: United Kingdom, United States, Australia, Austria, Belgium, Brazil, Canada, France, Germany, Ireland, Italy, New Zealand, Poland, Portugal, Singapore, South Africa, Spain, Switzerland, Turkey (plus Global regex fallback).
Verifiable Sanitization
Sensitive areas in PDFs are converted into flattened raster images. Text streams are sanitized, preventing reverse extraction from hidden layers. 100% Offline.
System & Accessibility
Requires Windows 10/11 (x64) and WebView2. Available in 8 UI Languages: English, French, German, Italian, Polish, Portuguese, Spanish, and Turkish.
How it works
1. Open a document
Drop in a PDF, scan or Office file. Everything is processed on your machine — no upload, no API calls.
2. Review the redactions
Doxudel proposes matches using deterministic rule packs plus on-device NER. Accept, reject or adjust each item in the calibration view.
3. Export sanitized output
Get a flattened, redacted file with no recoverable text layer, plus a redaction report you can attach to your DPIA.
The PII Redaction Landscape
The market forces a compromise between privacy, cost, and context accuracy. Here is how Doxudel's local NER architecture changes the baseline.
| Metric | Basic Regex Tools | Cloud & API LLMs | Heavy Local LLMs (70B+) | Doxudel (Local NER Architecture) |
|---|---|---|---|---|
| Context & Entity Detection | Fails (Rigid patterns only) | High | High (Hallucination risk) | High (Off-the-shelf NER + Deterministic Rules) |
| Data Privacy (GDPR/DPA) | Secured (Local processing) | 3rd-Party Transit Risk | Secured (Local processing) | Secured (Data never leaves device) |
| Chain-of-Custody | Maintained | Broken (External endpoints) | Maintained | Maintained |
| Infrastructure Cost | Very Low | High (API volume fees) | Massive (Enterprise GPUs required) | Low (Standard local hardware) |
| Stability (Hallucination) | Zero | High (Generative models drift) | High (Generative models drift) | No generative model (Deterministic extraction; fail-closed masking with published residual limits) |
Why Architecture Matters
The Flaw of Cloud Data Transit
Uploading sensitive files to third-party endpoints creates compliance and liability risks. Once data leaves your local machine, the chain-of-custody is broken. Furthermore, pay-as-you-go cloud redaction inflates operational costs based on volume.
The Hardware Burden of Local LLMs
Running local 70B+ LLMs solves the privacy issue but requires expensive enterprise GPU hardware. Generative models also suffer from semantic drift, offering unpredictable reliability for compliance operations.
The Doxudel Advantage
Doxudel pairs optimized, lightweight on-device NER (Named Entity Recognition) models with deterministic checksum validations. It runs entirely on standard local hardware without external API calls, providing predictable, rule-based outcomes.
Irreversible Raster Sanitization
Sensitive areas in documents are permanently flattened into sanitized rasters. This prevents reverse-extraction from hidden PDF text layers. The redaction is visually destructive and secure, leaving no residual metadata behind.