PRODUCT · PRIVACY-FIRST LOCAL AIv1.0 · Experimental privacy tool

RedactGuard

Detect sensitive information locally, review every finding, and move only the minimized document into the next workflow.

Local processingCustom PIIHuman review

See the workflow ↓

RedactGuard review screen showing detected sensitive data before redaction

Detect locally → review → redactNothing leaves automatically

The problem

Sensitive data is not a fixed list.

What must be protected changes by document, industry, organization and workflow. That makes detection both a privacy problem and a maintenance problem.

Manual review is slow

Long documents make anonymization repetitive, expensive and inconsistent.

Static rules are hard to adapt

New PII often means new regex, mappings, exceptions and application logic.

Cloud AI changes the privacy boundary

The original sensitive document must leave the machine before it can be analyzed.

The hidden cost: every change in what counts as sensitive can become a software change.

The solution

Define what is sensitive.
RedactGuard finds it locally.

Standard identifiers or domain-specific information can be expressed through profiles and custom PII instructions used by the local LLM.

For many taxonomy changes, you update the definition instead of building another hard-coded detector.

Define PIIUse a profile or add custom sensitive categories.

Detect locallyThe GGUF model analyzes the document on the machine.

ReviewThe model suggests. The user confirms.

ExportOnly the minimized result moves to the next workflow.

Product proof

Load. Review. Export.

The workflow stays visible and reversible until the user confirms the final redactions.

1 · LOAD

Choose the document and profile

Start from a built-in domain profile or a custom detection definition.

2 · REVIEW

Inspect every finding

Keep, remove or adjust what the local model identified as sensitive.

3 · EXPORT

Share the minimized result

Export only after the redaction choices have been explicitly reviewed.

Under the hood

Local by architecture.

Once the product idea is clear, the implementation is intentionally simple: UI, document processing, and local inference have separate responsibilities.

RedactGuard local-first technical architecture

React / Vite

Upload, profile selection, human review and export workflow.

FastAPI + Docling

Document conversion, sessions, PII orchestration, redaction and caching.

Local GGUF LLM

Contextual PII detection through llama-cpp-python, including profile and custom-type instructions.

Explore the architecture on GitHub ↗

⚠

RedactGuard is an experimental privacy tool, not a compliance guarantee. Local AI can miss, misclassify or over-detect sensitive information, so results still require human review.

EVIDENCE & LIMITS

The workflow is implemented. Detection is still fallible.

RedactGuard treats reviewability as part of the product boundary: local inference can accelerate detection, but it does not turn the model into an autonomous privacy or compliance authority.

EVIDENCE BOUNDARY

Local, configurable and reviewable does not mean perfect.

The current product supports local GGUF inference, configurable standard and custom PII definitions, reviewable findings and explicit export. The model can still miss, misclassify or over-detect sensitive information, so human review remains mandatory.

Supported today

Document analysis can run through local GGUF inference
Standard and custom PII definitions are configurable
Every finding remains visible and reviewable before export
The minimized result moves only after explicit user choices

Not claimed here

Compliance guarantee
Published perfect-recall or perfect-precision claim across PII domains

CONNECTED SYSTEM

Where this project fits

BUILD · INFRASTRUCTURELocal LLM ServerReusable local inference boundary

TEST · PRODUCTRedactGuardPrivacy-first proving ground

MEASURE · EVIDENCEPerformance LabRuntime viability on real hardware

Local-first data minimization

Define what is sensitive. Share less.

RedactGuard turns local AI into a configurable privacy workflow: define → detect → review → redact → export.

Explore RedactGuard ↗