Real data in throwaway prompts
Experimental prompts often contain the data the production system shouldn't send either.
Solution
Prototype AI faster without exposing sensitive data
Move fast on AI without leaking the test data.
Use Privian during experimentation so prototypes can work with real-shaped data without sending it to a provider. Same endpoint as production, scoped to test keys.
Problem
Prototyping is where data leaks start
Teams move quickly when prototyping AI: paste a customer email into a prompt, try a model, tweak it, try a different one. Synthetic data slows iteration; real data accelerates it but creates exposure. Most "small experiments" end up sending real identifiers and the occasional API key to a provider that was never approved for that purpose.
Why this matters
What's at risk during experimentation
Pasted secrets
Developers paste tokens into prompts to debug; those tokens then leave the network.
Untracked provider sprawl
Different teammates try different providers, each accumulating context.
Hard-to-unwind exposure
What was sent during prototyping cannot be retracted from the provider.
How Privian fits
Privacy-first experimentation
Fast to test
One POST to /v1/gateway. No SDK install, no rewrite.
Minimal integration
Set a base URL and a test API key. Keep your existing HTTP client.
Privacy-first gateway
Supported sensitive values are masked before any provider call.
Same shape as production
Prototypes behave the same way the production gateway will.
Architecture
Where Privian sits in a prototype
Notebook / script → Privian gateway → LLM provider
│ mask
│
(placeholders)
│
▼
rehydrate ← provider responseExample flow
Example request
curl https://api.privian.io/v1/gateway \
-H "Authorization: Bearer $PRIVIAN_API_KEY_TEST" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5.5",
"prompt": "Classify this support email from jane@acme.com.",
"stream": false
}'Swap the model string to compare providers. Swap the prompt to iterate. The data-handling story stays the same.
Use cases
Where prototyping with Privian pays off
Hackathons
Ship demos using real-shaped data without an after-the-fact review.
Internal AI experiments
Try ideas across teams without each one re-litigating data flow.
Prototype copilots
Build a vertical copilot prototype in a day, with masking on by default.
Proofs of concept
Validate an AI feature with customers before formal procurement.
Implementation
Developer integration path
Step 1
Add a provider
Plug an OpenAI, Anthropic, Google or DeepSeek key into the dashboard.
Step 2
Issue a test key
Use sk-gw_test_... in notebooks and prototypes; keep live keys for production.
Step 3
Iterate on prompts
POST to /v1/gateway with model + prompt + stream.
Step 4
Compare models
Swap provider/id to try different models against the same prompt.
Step 5
Graduate to production
Re-key with a live token; nothing else changes.
Transparency
Beta limitations
- No OpenAI SDK drop-in
- No native provider streaming (artificial chunking in beta)
- No tool / function calling support yet
- No per-request masking toggle
- Detection covers the supported entity set only
- No HIPAA / SOC 2 / PCI compliance claims today
FAQ
Frequently asked questions
- Can I use Privian during AI prototyping?
- Yes. Use a test gateway key against the same endpoint. Masking and rehydration behave identically, so prototypes can iterate on real-shaped data without exposing it to the provider.
- Does Privian require changing providers?
- No. Pick the model per request using a provider/id format such as openai/gpt-5.5 or anthropic/claude-sonnet-4.5. Switching providers is a string change, not a redesign.
- How quickly can teams integrate?
- A working request is one POST against https://api.privian.io/v1/gateway with model, prompt and an optional stream flag. Most prototypes are wired up in minutes once a provider key is added.
- Can I disable masking for a specific test?
- Masking is part of the security layer by design and is not toggled per request. Use a separate environment if you need to bypass it for a specific experiment.
- Is anything stored from my prototype calls?
- Raw prompts and responses are never persisted. Only structural counters (request, model, entity counts) are kept for observability and billing.
- Can I compare models safely?
- Yes. Send the same prompt to different provider/id values. Because masking happens at the gateway, swapping models does not change your data-handling posture.
Learn more