CPU-friendly local ML

Maskara

Ask AI without exposing people.

A compact PII detection model and SDK for adding privacy middleware to your codebase. It masks sensitive prompt data before cloud AI sees it.

Try the privacy pass View model

One local pass. No raw PII leaves.

Pick a realistic prompt, watch Maskara replace sensitive spans, then compare what the cloud sees with what your app restores locally.

Prompt composer Type or pick a sample with sensitive data

0/3,000

tiny browser demo

Privacy pass Ready.

Detect Find sensitive spans locally. 0 found

Replace Swap raw values with safe twins. local vault

Restore Put originals back after the AI reply. 0 sensitive spans

Cloud sees safe prompt

App receives restored reply

The public checkpoint lives on Hugging Face. The SDK is built to plug into an app before an LLM call.

One small wrapper around the prompt path.

1. ProtectDetect and mask local PII.

2. CallSend only safe context.

3. RestorePut originals back locally.

from maskara import Maskara

maskara = Maskara()
safe, ctx = maskara.protect(prompt)
reply = llm.complete(safe)
final = maskara.restore(reply, ctx)

Wrap a prompt, call your provider, restore the response. No new app architecture required.

Designed for local and tiny-server demos with deterministic rules plus a compact token classifier.

Names, emails, phones, addresses, credentials, and other prompt-level sensitive spans.

token classification local masking LLM middleware