A single engine for text, images, documents, audio and video.
PolyRedact exposes a unified redaction engine. Different content types all go through the same policy and logging layer, so you do not need five different tools for five different modalities.
End-to-end privacy layer
An end-to-end privacy layer for your AI data
Put PolyRedact across the full lifecycle: prep your data, gate live traffic, then hand auditors a tamper-evident trail.
Before training
Redact high-risk fields in tickets, chat logs, documents and transcripts.Do it before they are used for fine-tuning or retrieval.
During inference
Route live traffic through PolyRedact’s multimodal gateway.Do it before it reaches LLMs, analytics tools or external vendors.
Afterward
Keep signed audit trails and simple reports.Let security and privacy teams plug them into existing review and approval workflows.
Before training, During inference, Afterward — each phase gets a clear action so teams stay aligned on how data moves.
- Clean training corpora with automated redaction.
- Live gateway for screenshots, audio, video and text.
- Signed audit logs that slot into existing reviews.
Text redaction
Detect emails, phone numbers, addresses, IDs, credit cards, API keys, internal IDs and custom patterns. Return redacted text and a structured list of findings.Screenshot & image redaction
Handle screenshots of consoles and dashboards. Combine OCR and visual detection to find PII and secrets, apply blur or black boxes, and return safe images plus findings.Document redaction
Process PDFs and DOCX files by extracting text and embedded images. Run both text and image engines and produce per-page reports or fully redacted exports.Audio redaction
Transcribe audio with timestamps, detect PII in the transcript and map findings back to time ranges. Return redacted transcripts and lists of sensitive segments.Video redaction
Extract audio and key frames from video, run the audio and image engines, and generate a redaction report with preview frames – with the option to output fully redacted videos.Structured & database data redaction
Scan CSVs, SQL tables and JSON payloads to detect sensitive columns, mask PII consistently across rows, and return clean datasets plus a findings report.Gateway and API: plug into your existing flows.
Two modes: Proxy mode forwards to OpenAI, Azure OpenAI or your own models after redaction. Redaction-only mode lets you preprocess content and handle forwarding and storage yourself.
Before
const result = await openai.chat.completions.create({ ... });After (via PolyRedact)
const result = await polyredact.proxy({ target: 'openai', model: 'gpt-4o', messages, images });- Drop-in /proxy endpoint sits between your app and vendors.
- Redaction-only mode lets you keep forwarding under your control.
- Consistent audit logging across every outbound payload.
- Consistent policy model across text, screenshots, documents, audio and video.
One place to manage what is allowed to leave.
- Enable/disable categories like PII, secrets, internal IDs and custom regex patterns.
- Decide per category whether to block, mask, partially mask or allow but log.
- Use different policies per tenant or environment.
Every request generates an audit record with hashes of the original and redacted payloads, a summary of findings, timestamps and caller identity. Records can be signed with a service key so you can verify that logs have not been tampered with and export them to your SIEM.
FAQ
Text and screenshots are available now. Documents are in early access and audio/video are rolling out for design partners.
Yes. PolyRedact is built to run in your environment for teams that need to keep sensitive traffic inside their own cloud.
Use the /proxy endpoint to forward to OpenAI, Azure OpenAI or your own models after redaction, or call the redact-only APIs.
Share the signed audit trail approach and optional private cloud deployment.
View use cases