The API for Redaction
SecurePass is what you call before storing your data.
SecurePass is an API that redacts sensitive data before you store it. Phone calls, documents, email, images, and video go in. The same files come back with the values you should not be holding removed, so what you keep is ready to use.
A model does not need John's card number.
AI means you need data.
Models, evals, and benchmarks are only as good as the data behind them.
Data means you need compliance.
Call recordings, documents, and transcripts carry card numbers, Social Security numbers, account numbers, dates of birth, and addresses. Keeping them brings PCI, privacy, and retention obligations.
A model needs to know what a card number looks like and where one shows up in a conversation. It does not need a real one. It needs to know what a Social Security number is. It does not need John's. The specific values add risk and no signal.
Caller 00:41.2
Sure, it is 4111 1111 1111 1111, expires 04/29.Caller 00:41.2
Sure, it is [CARD_NUMBER], expires [CARD_EXPIRY].
CARD_NUMBER 00:41.9 to 00:47.6
CARD_EXPIRY 00:48.0 to 00:49.1Kept: where it was said, what kind it was, and how long it ran. Dropped: the digits.
Rarely is it that tidy.
Most card numbers arrive in pieces, with the agent talking in between and a correction partway through. SecurePass follows the digits across the interruptions and redacts every piece, while the agent's words stay.
Caller 01:12.4
Okay, it starts four one one one.
Agent 01:14.1
Mm-hm.
Caller 01:14.7
One one one one, one one one two.
Agent 01:17.9
Got it, go ahead.
Caller 01:18.6
Sorry, that last one was one one one one. Then one one one one.
Agent 01:23.0
And the expiration?
Caller 01:23.8
Oh four, twenty nine. Code is three two one.Caller 01:12.4
Okay, it starts [CARD_NUMBER].
Agent 01:14.1
Mm-hm.
Caller 01:14.7
[CARD_NUMBER].
Agent 01:17.9
Got it, go ahead.
Caller 01:18.6
Sorry, that last one was [CARD_NUMBER]. Then [CARD_NUMBER].
Agent 01:23.0
And the expiration?
Caller 01:23.8
[CARD_EXPIRY]. Code is [CARD_CVV].
CARD_NUMBER 01:13.2 to 01:14.0
CARD_NUMBER 01:14.7 to 01:17.4
CARD_NUMBER 01:19.8 to 01:21.1
CARD_NUMBER 01:21.6 to 01:22.8
CARD_EXPIRY 01:23.8 to 01:25.0
CARD_CVV 01:25.6 to 01:26.5Four pieces, one correction, and the agent's replies in between. Every piece is redacted, including the group the caller corrected, and only the digits are gone. The same happens when the number is typed on the keypad, where the tones are found and silenced.
Any file you were about to store.
The same intake as FirstPass. A file SecurePass cannot read is kept and flagged for a reviewer, never dropped.
Call recordings
- WAV
- mu-law
- MP3
- M4A
- Opus
Single or dual channel, with keypad tones.
- MSG
- EML
- ICS
Threads parsed, attachments expanded, messages nested inside messages included.
Office
- DOCX
- XLSX
- PPTX
- DOC
- XLS
- PPT
- RTF
Current and legacy formats.
OpenDocument and Apple
- ODT
- ODS
- ODP
- Pages
- Scanned PDF
Scanned pages are read visually, not by character recognition.
Text
- TXT
- CSV
- Markdown
- Log
- HTML
Images
- JPEG
- PNG
- TIFF
- GIF
- BMP
- WebP
EXIF and GPS data read and masked too.
Audio and video
- MP4
- MOV
- WebM
- MP3
- WAV
Transcribed, then masked as text with timestamps.
Archives
- ZIP
Expanded, and each item handled on its own.
Three steps from raw to ready.
Each redaction lines up with the moment it was said, in the audio and in the transcript.
Send a recording or a file, or point us at your bucket
Post a single call or file to the API. Or give SecurePass access to the folder in your S3 bucket where your recordings and files already live.
We transcribe and find the sensitive spans
Each call is transcribed with word timings. Sensitive spans are found in the words and in keypad tones, then mapped to exact times in the audio.
You get redacted audio and a redacted transcript
The audio comes back with those spans silenced. The transcript marks each one, with the start and end time of every redaction.
Recordings in your own bucket are read with a role you grant, and the results are written back to your bucket. Nothing needs to be uploaded or downloaded. Recordings sent to the API are deleted after processing unless you choose otherwise, and redacted results are kept for 30 days.
Data you are allowed to keep is data you can use.
Masked calls and documents are a clean base for measuring models today, and for building on later.
Evaluate
Build evaluation sets from real calls and documents. The structure of each interaction stays. The real values do not.
Benchmark
Create custom benchmarks on your own call and document types, so you measure models on the work they will actually do.
Personalize
Generate realistic stand-ins with the same shape: a card number read in four groups with "uh huh" between them, a mistyped account number, a hold mid-call. Then tune models on data you are allowed to keep.
One call, a whole bucket, or a step in an agent.
REST API
Send one recording. Get back redacted audio and a redacted transcript.
curl -H "authorization: $KEY" \
-F file=@call.wav \
https://api.autessa.com/clients/call-redaction/redactCommand-line client
Run bulk jobs over a bucket of recordings with callredact, and follow them as they finish.
callredact submit-s3 s3://bucket/recordings/ \
--output s3://bucket/redacted/
callredact watchAutessa agent tools
Use the same building blocks inside Autessa agents and scripts.
One cent per minute of audio.
- Redacted audio, redacted transcript with word timestamps, and the time range of every redaction, all included.
- Billed on the length of the recording. No minimums and no setup fee.
- Single calls and bulk jobs over a bucket are priced the same.
- For volumes in the millions of minutes, talk to us.
Redact before you store. Then you are ready to go.
Start with your call recordings or a folder of files. Send a single item, or point SecurePass at a bucket.