Private AI architecture

AI document tools that keep document inference off third-party AI model APIs.

On VersePDF.com, uploaded document content and AI prompts are processed within VersePDF-controlled infrastructure. Production is configured fail-closed: OpenAI, Anthropic, Gemini, and other third-party AI model APIs are not used for document AI inference.

No third-party AI model API

Document AI content stays on the VersePDF inference stack.

Machine-verifiable posture

The public /v1/privacy/ai endpoint exposes whether external AI model APIs are enabled—never the keys.

Designed for sensitive work

A stronger starting point for legal, healthcare, finance, and compliance reviews—without claiming certifications we do not hold.

What “private AI” means here

Browser upload → protected VersePDF processing → result
Document reasoning & generation
VersePDF-controlled private reasoning service
Document vision
Donut and TrOCR models inside the VersePDF AI container
Speech transcription
Self-hosted Whisper
Document translation
Self-hosted M2M100 translation service
Image captioning
Self-hosted BLIP
Text to speech
Self-hosted SpeechT5
Image processing
Local rembg / ESPCN processing and browser-local tools where applicable
Grammar analysis
Self-hosted LanguageTool plus local Hunspell rules

Model packages may be downloaded from their publishers during image build or model provisioning. That is different from sending a customer's document or prompt to the model publisher. Runtime document inference is performed on the self-hosted services above.

What this claim does—and does not—cover

The precise claim is about AI inference on uploaded document content and prompts. VersePDF is an internet service, not an air-gapped appliance, and several optional non-AI integrations have separate data flows. Those boundaries are disclosed instead of being hidden behind a vague “private” label.

Payments
Paystack receives the payment information required to process a purchase; document AI content is not sent as part of payment processing.
Google / Apple sign-in
OAuth providers receive the normal sign-in exchange, not document AI prompts or uploaded files.
Canva export
Only runs when you explicitly export to Canva; the selected export is then sent to Canva.
YouTube transcript
The requested public video ID/caption request is sent to YouTube's caption service; this is not uploaded-document AI inference.
Web-search tools
Plagiarism/fact-check features that require live web search can send selected search phrases or claims to the configured search provider.
Public timestamping
Cryptographic timestamping can contact an external timestamp authority as part of the signing workflow.
Diagnostics
Sentry can receive scrubbed technical error telemetry when configured. Request bodies, cookies, and authentication headers are excluded; strict content-egress deployments disable Sentry entirely.

Privacy architecture is not a compliance certificate

Self-hosted inference is a technical control. It does not by itself make a service SOC 2 certified, ISO 27001 certified, HIPAA compliant, GDPR compliant, or suitable for every regulated workload. VersePDF does not claim a certification unless it has been formally completed and documented. Enterprise buyers should review retention, access controls, subprocessors, incident response, backups, contracts, deployment architecture, and their own regulatory requirements.

Evaluating private document AI for your organization?

See the enterprise review path or contact us for a technical discussion.