@devadit15

Improved eye tracking and added some instructions to make it easier too use. Disabled mouse control over the highlighted box as it confuses users. Also fixed the bug in the audio during the model training.

Everytime you paste something into an AI you hand over everything in it — your name, account number, the medical stuff you’d rather nobody saw — and the model doesn’t actually need any of that to be useful, it just needs the shape and relationships of the data, not the real identities. So I built Cloakroom as a passion project, nobody asked me to, I just got obsessed and wanted to see if I could pull it off end to end. It’s a privacy layer that sits between your app and any AI and runs a four-step loop: detect the sensitive bits, mask each one with a stable token like [PFI_ACCOUNT_1], let the model reason over just the tokens, then put the real values back in the final answer so only you see them. The stable part matters — the same value always maps to the same token, so a name that shows up three times stays one token all three times, meaning the model can still reason about who owns what without ever learning who’s who. A few things got tricky: a 12-digit number in India could be a bank account or half an Aadhaar, so I use the surrounding fields to tell them apart, and a “person” in a financial context isn’t just a person, they’re the account holder, a different sensitivity class. I also scan the model’s output afterward in case it leaks or hallucinates a real-looking number on its own, and flag it. And it only logs counts, never raw values, so you can prove what got masked without keeping a second copy of the exact data you were protecting. It grew a few faces too — a Python library, an HTTP service any language can call, a live web playground where you paste a record and watch it get masked and restored in real time, and a browser extension for the “just protect me as I type” version. It’s also cloud-agnostic: the detector, the encrypted vault, and the model are all swappable with one setting, so it runs completely free with zero paid keys out of the box, but flip a couple env vars and the same code runs on managed cloud services at scale, no rewrite. Not everything’s done — free-text names need a real NER model over plain regex, a couple cloud adapters are half-built, and there’s a tuning dashboard I sketched but haven’t made, though the audit stream already emits the data it’d need. Next I want to finish that dashboard, round out the adapters so it’s genuinely any-provider, push detection smarter with better NER and more region-specific ID formats since India was just the start, and make the extension seamless enough that you forget it’s there — because that’s the whole point, privacy shouldn’t be something you remember to switch on, it should just be there by default like HTTPS, and Cloakroom is my attempt at making that normal.

Improved eye tracking and added some instructions to make it easier too use. Disabled mouse control over the highlighted box as it confuses users. Also fixed the bug in the audio during the model training.

UI Redesign

UI Redesign