墨 Sumi — Japanese PII detection

A 0.13B token classifier plus a rule layer, running entirely in your browser. Nothing you paste is uploaded anywhere — the model is downloaded to your machine and inference happens locally. For a redaction tool, that seemed like the right way to build the demo.

This does not guarantee legal or regulatory compliance. It reduces the risk of personal data leaving your machine; it is not a complete detector, and misses will happen. All training and evaluation data is synthetic — no real personal information was used in this project.

Input (Japanese)

Loading the model (about 133 MB, cached after the first visit)…

Detected — compared with Presidio + GiNZA

both agree Sumi only Presidio + GiNZA only
Results will appear here.

Redacted output — this is what you would send to an LLM


      

        
        
      

    

    

      
Mapping table — original values are never shown

      

        PlaceholderTypeMaskedLenOffsets
      

      

        In the library this table is written to a local map.json with mode
        0600 and never transmitted. If an original value would ever reach
        the wire, the send raises rather than logs.