README.md

---
title: Redac
emoji: 🖍️
colorFrom: gray
colorTo: red
sdk: gradio
sdk_version: 5.44.1
python_version: '3.13'
app_file: app.py
pinned: false
license: other
short_description: Redac
tags:
- track:backyard
- sponsor:openai
- sponsor:openbmb
- achievement:offgrid
models:
- openbmb/MiniCPM-V-4_5
- urchade/gliner_multi_pii-v1
---

# Redac

Redac is a local privacy gateway: paste sensitive text or upload a document/ID
image, extract and redact PII locally, then copy the safe output into any LLM
without exposing raw personal data. The local placeholder map lets you rehydrate
answers afterward.

## Submission

- **Demo video:** https://youtu.be/as7GmNYO83s
- **Social post:** https://x.com/barathishere/status/2066882507497242782

## Build Small Hackathon Badges

| Badge | Why Redac qualifies |
|---|---|
| 🏡 **Backyard AI** (`track:backyard`) | Redac solves a practical privacy problem for people who want to use LLMs with medical, ID, banking, or personal documents without leaking raw PII. |
| 🌀 **OpenAI Track** (`sponsor:openai`) | The Space history includes Codex-attributed commits, and the app is designed around a copy-safe workflow for downstream LLM use. |
| 🏮 **OpenBMB Awards** (`sponsor:openbmb`) | Image extraction uses `openbmb/MiniCPM-V-4_5` in-Space via ZeroGPU before the extracted fields flow into the redaction core. |
| 🔌 **Off the Grid** (`achievement:offgrid`) | Extraction, detection, redaction, and rehydration run inside the Space process. No external AI API receives raw PII. |

## Models

- `openbmb/MiniCPM-V-4_5` — local document/ID image extraction.
- `urchade/gliner_multi_pii-v1` — local PII-tuned zero-shot NER.

## Flow

1. Paste text or upload a document image.
2. Redac extracts fields locally when needed.
3. Redac replaces detected PII with stable placeholders.
4. Copy the safe output into your LLM.
5. Rehydrate the LLM answer locally when you need the original values back.