Chat with cloud LLMs while a local model swaps your PII out before the prompt leaves your machine — and rehydrates the answer back to your real entities.

Every time you ask a cloud AI about something personal — a city you're moving to, a company you work for, a name — you're sending that data to an external server. Cloak intercepts each prompt, runs a local PII detector, replaces sensitive entities with realistic stand-ins (not [REDACTED] — actual fake names and cities), sends the scrubbed prompt to the cloud, then maps the answer back to your real entities before you see it.

The round-trip substitution is the differentiator: the cloud model sees a coherent, answerable prompt; you get a useful, real answer; and the actual PII never left your device.

This is disclosure control, not a privacy guarantee. The pipeline reduces PII exposure significantly — but a determined adversary with enough queries could still reconstruct context. The goal is to make routine LLM use substantially less leaky, without sacrificing answer quality.

Prerequisites: Ollama installed and running.

# 1. Install Ollama (if not already):
#    https://ollama.com/download

# 2. Clone and launch:
git clone https://github.com/praxis-society/praxis-cloak.git
cd praxis-cloak
./scripts/run.sh

run.sh will:

• Download praxis/spanfinder-3b and praxis/relevance-3b from HuggingFace and build them into Ollama locally (first run only, ~3.8 GB total — no Ollama account needed)
• Create a Python virtual environment and install dependencies
• Build the frontend (requires Node; skipped if npm is not found)
• Start the local server at http://127.0.0.1:8765 and open it in your browser

Settings: Paste your Anthropic or OpenAI API key in Settings to route scrubbed prompts through your preferred provider.

run.sh downloads these GGUFs from HuggingFace and builds them into Ollama for you (the relevance model ships a chat template via its Modelfile, so build it locally rather than pulling the bare GGUF). Both are Qwen2.5-3B-Instruct LoRA fine-tunes; q8_0 reference quantizations (~3.1 GB each) and the LoRA adapters are also on HuggingFace. To try the span detector standalone: ollama run hf.co/praxis-nation/spanfinder-3b:Q4_K_M.

User prompt
    │
    ▼
[regex fast-pass]          ← catches obvious patterns (email, phone)
    │
    ▼
[spanfinder-3b]            ← on-device; detects PII spans (names, orgs, locations)
    │
    ▼
[relevance-3b]             ← decides: KEEP (load-bearing) or SCRUB (incidental)
    │
    ▼
[substitution]             ← replaces SCRUB entities with realistic fakes
    │                         (not [REDACTED]; actual coherent stand-ins)
    ▼
[cloud LLM]                ← sees a coherent, answerable, scrubbed prompt
    │
    ▼
[rehydration]              ← maps fake entities in the answer back to real ones
    │
    ▼
Answer (real entities)

Re relevance step - Keyword scrubbing over-removes: "What's the tax rate in Toronto?" needs "Toronto" to produce a useful answer. The relevance model judges whether each detected entity is needed — and only the incidental ones get substituted.

Re realistic fakes - [REDACTED] breaks the cloud model's ability to reason coherently. A fake city name preserves sentence structure and answer quality; the rehydration step maps it back.

app/
  backend/          FastAPI server (scrub pipeline, cloud calls, rehydration)
  frontend/         React UI
models/
  spanfinder-3b/    HuggingFace model card
  relevance-3b/     HuggingFace model card
  Modelfile.spanfinder
  Modelfile.relevance
  LICENSE_AND_ATTRIBUTION.md
scripts/
  run.sh            One-command launcher
  install.sh        Dependency installer (no server start)
src/                Core library (span detection, substitution, pipeline)

Source code: MIT. Copyright (c) 2026 Praxis. See LICENSE.

Model artifacts (*.gguf, *.safetensors, LoRA adapter weights): Qwen Research License (non-commercial). Commercial use requires a separate license from Alibaba Cloud.

Built with Qwen — praxis/spanfinder-3b and praxis/relevance-3b are fine-tuned derivatives of Qwen/Qwen2.5-3B-Instruct. Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved.

See NOTICE and models/LICENSE_AND_ATTRIBUTION.md for full attribution including the CAPID dataset (MIT) used to train the relevance model.