WordZilla runs an AI model on your computer to find every name, address, email, phone number and account number in your documents, and replaces them with stable coded tokens. Only the redacted copy is ever sent for analysis — and the real names come back on your machine alone.
From an empty install to a finished analysis — the four moments that matter.
WordZilla — Super Easy ModeOn 14 March 2024, Sarah Loveridge of Harbor Point Properties LLC wrote to counsel for Dalton & Reyes regarding the property at 1420 Ashcroft Avenue, Suite 300, Fullerton, CA 92833, confirming that the parcel identified as TMK No. (1) 9-6-008-007 had passed inspection. Correspondence should be directed to s.loveridge@harborpointllc.example or (808) 555-0142 during the review period.
No pipeline to configure, no model to install by hand, no cloud account needed to redact.
PDFs (scanned ones are OCR'd), Word, Excel, PowerPoint, email, plain text. WordZilla creates its work folder next to them and leaves your originals untouched.
A model running on your own machine reads every document and replaces identifying details with coded tokens like «PER_0001». The same person gets the same token in every file.
You see the redacted corpus and can add anything the model missed, then re-run. Nothing is sent anywhere until you explicitly click Start.
Send it to Claude or GPT with your own API key — or keep it entirely offline and analyze with the local model instead.
The redacted analysis, plus a copy with the real names restored from your keypair. The restoration happens on your computer.
Chat with Docs answers questions about your files with the local model, citing its sources — click a citation and the original document opens on the exact page, highlighted.
Redaction is local because it has to see the sensitive material. Analysis is remote because that's where the reasoning is. Only redacted text crosses over.
No detector catches everything, and WordZilla is built on the assumption that it will miss things. That's why it shows you the redacted corpus before anything is sent, lets you add what it missed and re-run, and re-scans its own output for surviving details. You remain responsible for verifying the redacted text before you share it. The app says the same thing at the point where it matters.
Not a stripped trial: it redacts real documents, runs real analyses, and produces real output you keep.
Both guided modes: Super Easy Mode (pick a folder, get redacted documents and a cloud analysis) and Chat with Docs (ask questions about your files locally, with clickable citations into the source pages). Redaction, review, re-run, local or cloud analysis, and automatic unredaction of the results.
A GPU makes this pleasant; on CPU alone a large corpus can take hours, and the app tells you so up front with a time estimate before you commit. The demo runs one specific local model so behaviour is predictable. Your redacted output and keypair stay yours when the demo ends.
Windows 10 / 11. No account and no API key to start. The first run downloads the ~4.2 GB private AI model to your own machine; after that it works offline.
This build is not yet code-signed. Windows SmartScreen will show “Windows protected your PC / unknown publisher” the first time you run it — choose More info → Run anyway. A signed build is coming; until then, only install it if you are comfortable with that prompt.
Leave an email and we'll send one message the day it's ready — no drip campaign. If you tell us what you work on, we'll prioritise accordingly.
We store your email address to send you the release notice, and nothing else. No tracking pixels, no third-party analytics on this page. Ask us to delete it any time.
No. Ingestion, OCR, PII detection and redaction all run on your machine, using a model that also runs on your machine. The only thing that can ever leave is the redacted corpus, only when you click Start Cloud Analysis, and only to the provider whose API key you entered. You can skip that step entirely and analyze locally.
Not to redact — that's the whole local half of the app, and it needs neither. A key from Anthropic or OpenAI is only needed if you want a frontier model to do the analysis. The demo ships with no key of ours; you use your own, and the request goes straight from your machine to the provider.
Good enough that reviewing it is a review, not a rewrite — but not perfect, and we don't claim it is. It reads every document with a language model rather than pattern-matching, handles the same person appearing as "Sarah Loveridge", "Ms. Loveridge" and "S. Loveridge", survives OCR damage and line wrapping, then re-scans its own output for anything that got through. Then it shows you the result and asks you to check it. Verifying before you share is your call and your responsibility.
Yes — that's the point of coded tokens rather than black boxes. Your work folder holds a keypair that maps every token to its original value, and the app automatically produces an unredacted copy of any analysis alongside the redacted one. That restoration happens on your computer; the keypair never leaves it.
Any modern Windows PC will run it. A GPU (AMD, NVIDIA or Intel — it ships a Vulkan build) makes it dramatically faster; without one it falls back to CPU, which works but can take hours on a large corpus. The app estimates the run time before you start and warns you when an overnight run is on the cards.
The app stops and offers a $9.99/month subscription or a code. Everything it already produced — your redacted documents, keypair and analyses — is yours and stays on your disk.
Not yet. The codebase runs on Linux, but the packaged, signed demo is Windows-only for now. Join the list and say which you'd want — that's how we'll decide what comes next.