CamoText 🛡

Setting up client PII anonymization in your Claude Cowork process is easier than you think. Claude works just as well, if not better, with explicitly labeled subjects. It’s possible to preserve both output quality and confidentiality! camotext.ai/blogposts/use-…

@jan__kubica @BitGrateful @Law_Flows In our case, in each location the original term is swapped back verbatim for its corresponding anonymized tag (via the local key). The anonymization can have both user-defined custom-tagged terms, and auto-detected terms. Our international app also supports ES, DE, FR, and IT!

thanks for the context, makes sense my complaint about de-anonymizing was that in some languages, the word changes based on the sentence it is in, so it effectively has many variants. You catch the one mentioned by the user, but the issue is then swapping for another possible variant that fits the LLM answer. But I appreciate this is not much of a problem in English

Any lawyers using something like this? How does it affect productivity?

@jan__kubica @BitGrateful @Law_Flows We use a combination of local NLP models, rule & pattern-based, and other algos. Locally reinserting original terms/de-anonymizing using local keys also an app feature! x.com/CamoText/statu…

No idea what they specifically use. Usually you either run a local model (slow), trust a third party to anonymize for you (defeats the purpose), or rely on rule-based anonymization (lower accuracy). Ultimately as @Law_Flows mentioned, no approach is 100% effective. I put together an anonymization engine for my own product (github.com/stella/anonymi…), and the experience made me even more reserved about using it without a human pass over the text. Even if everything works correctly, swapping the original names back into the AI's output produces grammatically broken text in those languages where you need declined form of a word, not its base form. Less of a concern in English, sure, but for me personally another pain point ...

@Law_Flows @BitGrateful Exactly right: not only is the software imperfect because it's small enough to run locally on laptops, but privacy and confidentiality can be heavily subjective. Need a human in the loop.

@BitGrateful None of these PII stripping services are 100% effective. You still need a human review step at the end, but it can definitely save a lot of time by completing the first 95% or so. It's a significant time saver over the manual process.

@BitGrateful @lower_case_b In part. We use fully local NLP + custom algos, smaller and more performant on laptops.

@CamoText @lower_case_b local AI to scrub it?

Lawyers, do not use Fable for legal work. Fable 5 requires mandatory 30-day retention and has possible human review. Non-starters

@lower_case_b @BitGrateful ^ this right here folks

@BitGrateful Just PII scrub your data on the front end then no zdr issues

FYI if you have a zdr with anthropic. It doesn't apply to mythos/fable support.claude.com/en/articles/15…

We’re building a suite of multilingual privacy tools: camotext.ai/blogposts/mult…

Using AI with confidential, multilingual documents? CamoText International anonymizes English, Spanish, French, German, and Italian text with native models and interfaces. Switch with a click or CLI arg. Still fully offline, performant on laptops.

No usage throttling, unlimited use regardless of WiFi, no worrying if a site downloads your output or malware. Convert files to friendly/lighter formats, extract audio from video, voice typing, redacting - you don't need an internet connection. camotext.ai/camosuite/

The CamoSuite apps are designed for a maximal privacy by default (fully offline operation, nothing leaves your device). Also, some daily tasks are just better suited for local apps than cloud services, regardless of privacy:

Contractual privacy isn't adequate for professionals using AI. DPAs and BAAs are paper promises; often still inadequate for the confidentiality requirements of some professions. ZDR addendums are expensive, and still rely on trust. Why rely solely on contracts?

Great for privately converting: - PDF or .docx into Markdown, for AI to read more efficiently - meeting or deposition videos into an MP3 or WAV for transcription - HEIC iPhone photos to web-friendly WEBP and more: camotext.ai/camoconvert

File conversions can be a hassle; it's tempting to upload to an online converter and brave a maze of ads and questionable links. There's a more private way. CamoConvert is a desktop app that converts video, audio, images, and docs fully locally and offline.

Chatrie #legalizeprivacy

@ErikVoorhees + add our prompt/context doc anonymizer!

Verifiable round-trip privacy on a frontier agentic open-source LLM

@solirvine Definitely a challenge, but there are some contextual keys to help mitigate

Challenges with redaction/pseudo for legal contracts.

@zenpanda01 @gajesh Exactly, surprisingly few categories and half are catchable with pattern matching. Kind of validating for us

@gajesh How better is it compared to presidio ? We don’t necessarily need an LLM to detect PII

OAI cooked here. 1B model to redact all PII at client side.

CamoVoice, our offline speech-to-text desktop app, now has voice typing. Hit the hotkey, speak, hit again, and text is written at your cursor. Works in any app with no cloud, no subscription, no always-on listener: fully private. Audio never leaves your device. More info: