REDACT

Everything runs locally in your browser — files are never sent anywhere.

• Word docs are edited at the XML level: matched text is overwritten with block characters directly in the document, so it's genuinely removed, not just covered up. Only the main body is covered — headers, footers and footnotes aren't scanned yet.
• PowerPoint slides are redacted the same way — text runs are overwritten directly in each slide's XML. Speaker notes aren't scanned yet.
• Excel workbooks have their text cells redacted directly in the workbook XML. Numbers stored as actual numeric values (not text) are left alone, since altering them could silently break formulas elsewhere in the sheet — only text-type cells are covered.
• PDFs are flattened into images with black boxes drawn over matches, then rebuilt into a new PDF. This is deliberate: it guarantees no selectable text remains hiding under the black box, at the cost of the output no longer being text-searchable.
• Images are scanned with on-device OCR (Tesseract) to find matching words, which are then blacked out on the image.
• CSV, TXT, JSON and other text-based files are redacted directly as plain text — simplest and most reliable path, since there's no formatting to preserve.
• Anything else is treated as plain text too, as a best-effort fallback. Genuinely binary formats (video, audio, compressed archives, etc.) can't be safely redacted this way and will show an error instead of a corrupted file.
• Name detection is a best-effort NLP model (Compromise.js), not a guarantee — always review the result before sharing a sensitive document.
• Matching is case-insensitive and whole-word, so partial coincidental matches inside other words are avoided.
• Names to never redact always win — if a name appears there, it's left alone even if it's also typed into the redact list or found by auto-detect.