Full document redaction with Qwen 3.6 27B with a Pi agent harness

The author demonstrates that local models, specifically Qwen 3.6 27B, can perform end-to-end document redaction when optimized with a higher quantization level and an agentic harness using the PI framework.

• The system uses Qwen 3.6 27B quantized to Q6_K_XL with a 114k token context window, requiring 40GB VRAM.
• A Gradio-based UI allows users to upload documents and provide custom redaction instructions to the agent.
• The backend employs OCR tools like Tesseract and PaddleOCR alongside spaCy for PII identification.
• VLM capabilities are utilized for detecting faces and signatures within scanned documents.
• Testing on emails, partnership agreements, and government policies showed acceptable initial redaction results.

This approach offers a significant time-saving alternative for users needing to perform contextual-aware redaction tasks locally without relying on proprietary cloud APIs.