Sensitive information classification
Every request is labelled Public, Internal, Confidential or Restricted from what it contains, so policies and alerts can follow the level.
Data protection for authorised AI applications
PrivacyLayer is a proposed data protection layer that sits between your application and the model. It finds personal and confidential information in prompts and responses and applies your organisation’s own masking and redaction policy — before anything leaves your control.
Refund the charge for Adaeze Okafor, card 4111 1111 1111 1111, email adaeze.okafor@example.com.
Refund the charge for , card , email .
01 · Live playground
This runs the prototype’s detectors in your browser — nothing is sent anywhere. Edit the synthetic text or change what happens to each type of data.
Hi, I'm Adaeze Okafor (a••••••••••••@example.com, •••• ••• 0142). My card •••• •••• •••• 1111 was charged twice on order [ID]. Please refund to account ••••••6789. I was diagnosed with [HEALTH] last year, so the pharmacy delivery matters. My ID number is [ID] and my date of birth is [DOB]. Test key: [SECRET].
10 detections · 9 protected
02 · Core features
Every request is labelled Public, Internal, Confidential or Restricted from what it contains, so policies and alerts can follow the level.
Emails, phone numbers, Luhn-checked cards, account and ID numbers, dates of birth, names, health terms and credentials — each with a confidence.
Mask (keep the last four), redact to a label, hash to a stable pseudonym or allow — chosen per entity type.
Watch what flows into prompts and out of responses, by application, so leaks in model output are caught too.
Start from a fintech, healthcare or analytics profile and tune it; restricted data can only be allowed with a written reason.
Counts, actions and levels for every call — metadata and a fingerprint only, never the raw text.
One authenticated call before the model and one after. Same policy, same log, any provider.
03 · The differentiator
A configurable data protection layer built specifically for authorised AI application inputs and outputs. Prompts are cleaned before they reach the model; responses are checked before they reach the user.
“My card 4111 1111 1111 1111 was charged twice.”
→ “My card •••• •••• •••• 1111 was charged twice.”
“I’ll refund it to account 0123456789.”
← “I’ll refund it to account ••••••6789.”
04 · Who it’s for
Add a privacy layer to every model call without rebuilding your stack.
Keep card, account and identity numbers out of prompts and logs.
Pseudonymise patients and health details before summarisation.
Decision-makers CTOs Data protection officers Enterprise security managers
05 · Business model
Usage-based subscriptions tied to processed data volume, with enterprise functionality — custom detectors, single sign-on, customer-managed keys and dedicated support — in a higher tier. Illustrative pricing for a proposed product.
06 · Planned AWS infrastructure
PrivacyLayer is a proposed data protection layer for AI applications run by AI SaaS developers, fintech companies and healthcare technology providers. It must inspect text on its way into and out of AI models in milliseconds, apply each organisation’s own policy, and keep tamper-evident metadata without storing sensitive content. Designing on AWS lets it use managed PII detection and guardrails behind an authenticated API, with keys under customer control.
Amazon API Gateway
Applications send prompts and model responses to an authenticated privacy API.
Amazon Comprehend
Personal information is identified and located in the text, alongside custom detectors for each organisation.
AWS Lambda
Each entity is masked, redacted, hashed or allowed according to the organisation’s policy.
Amazon Bedrock Guardrails
Configurable protections filter sensitive content in inputs and outputs of Bedrock-hosted models.
Amazon DynamoDBAWS KMS
Policies and audit metadata are stored, encrypted with customer-managed keys; raw text is never kept.
PrivacyLayer is a proposed AI data protection platform for organisations developing and operating AI applications. The planned product will identify sensitive information, apply configurable redaction policies and support privacy monitoring within authorised application workflows. Amazon Comprehend will assist with personal information detection, and Amazon Bedrock Guardrails will support configurable application protections. We plan to use Amazon API Gateway and AWS Lambda for processing, Amazon DynamoDB for policy management and audit metadata, and AWS KMS for encryption. Organisations will configure their own privacy requirements and review detection accuracy before production deployment.
Planned, not yet deployed. This website and the privacy console run in your browser on synthetic text.
Founder
Elizabeth founded PrivacyLayer after seeing product teams pause promising AI features because nobody could say what personal data was flowing into prompts and out of responses. Elizabeth leads product, customers and every pilot, with a focus on configurable, measurable protection.
07 · Progress & demos
PrivacyLayer is a proposed product in its building stage. A demo walks through detection, policies, output checks and review on synthetic text. The MVP covers synthetic text analysis, configurable redaction rules, privacy review reports and an authenticated data protection API.
By Elizabeth Chinedu, to let teams use AI without sending personal data where it shouldn’t go.
Detect, classify, apply per-entity policies on inputs and outputs, log metadata only, review accuracy before production.
Live redaction, configurable policies, false-positive review, dashboards, audit log and API examples on synthetic text.
PII detection, guardrails, processing API, policy store and key management designed.
An authenticated data protection API in front of one authorised AI workflow, with accuracy reviewed on the customer’s synthetic data first.
FAQ
No. The audit log keeps counts, actions, the sensitivity level and a fingerprint of the protected output — never the raw text.
The prototype detects emails, phone numbers, Luhn-checked payment cards, bank account and ID numbers, dates of birth, names from a list, health terms and API keys. The planned product adds Amazon Comprehend and organisation-specific detectors.
Accuracy depends on your data, which is why every organisation reviews detections on synthetic data first and feeds false positives back into its policy.
Not yet. It is a proposed product; the privacy console is a prototype running in your browser on synthetic text.
Contact
Questions about the product, the API and integrations, security, or partnerships. Messages go directly to the founder.