Nvisy is in private beta. You cannot sign up for it, which is deliberate: we would rather work closely with a handful of teams than open the door and lose track of what breaks.
Why a closed beta
Redaction is unusual in that the failure is silent. A tool that misses a social security number does not throw an error or crash a page; it hands back a document that looks finished. The only way to find out is for someone to read the output and notice.
That is not something we can learn from usage graphs. It needs conversations with the people who run the documents through, about the formats we have not seen, the entity types that matter in their industry, and the redaction that looked right but was not. Those conversations do not scale, which is why the group is small.
What works today
The engine is the part we are confident in.
Detection reads over fifty entity types across personal, health and payment data, combining pattern libraries that carry real checksum validators with models that weigh the context around a match. Every match carries a confidence score, and you set the floor.
Redaction is permanent and byte-level: a field can be blacked out, masked, hashed, tokenized, pseudonymized, encrypted reversibly, truncated or generalized, and the same document can be redacted again under a different policy. For PDFs the metadata and hidden layers go too, which is where a redacted file usually leaks.
Around that: policies shaped to GDPR, HIPAA and PCI; review in threads with assignment and verification; an audit trail behind every detection; a documented REST API with signed webhooks; and connections to Drive, Dropbox, OneDrive, Box, S3, Azure and Google Cloud Storage.
Most of it is open source. The detection framework, the pipeline and the server are Apache-2.0, the console is AGPL-3.0, and the clients are MIT, so you can read how a document is handled before you hand one over.
What is still rough
It is more useful to say so than to let you find out.
The console has gaps where a workflow assumes you already know what you are doing. Error messages are not always specific enough to act on. Some formats are handled well and others are handled adequately. Throughput on large batches is fine but not tuned. There is no self-service billing, which is part of why there is no sign-up button.
None of that is a surprise to us, and none of it is dangerous to a document. But it is the difference between a product you evaluate and a product you deploy without thinking, and we are still in the first category.
Who it suits
The beta fits teams who have a concrete redaction problem rather than a general interest: a disclosure process that runs on people reading pages, a dataset that cannot go to a vendor as it stands, a pipeline that needs sensitive fields gone before a model ever sees them.
It suits you less if you need a finished product this quarter, or if you cannot spare the time to tell us when something is wrong. The feedback is the point.
Asking for access
Write to [email protected] or use the contact page. Tell us what you are trying to redact, roughly what volume, and whether you need it self-hosted. That is enough for us to say whether the beta is a fit, and we would rather say no early than waste your time.
If you would rather look before you ask, the source is public and the documentation is written for someone deploying it, not for someone being sold it.