The product
One governed surface between your team and every AI model
Wardary wraps cloud AI in the controls a regulated team needs: protect the data, govern the models, record every prompt — without slowing your people down.
The life of a prompt
What happens between “send” and the model
Four controls run on every message — inbound and outbound — so nothing sensitive slips through.
Draft a demand letter for our client [NAME_1] regarding matter [MATTER_2]. Their SSN [SSN_1] is on file.
Sent via encrypted tunnel · sensitive spans never left your boundary
I've drafted the demand letter for your client. Real names and identifiers were restored only here in your view — the provider saw placeholders.
- 01
Inspect & detect
The prompt and any attachments are scanned for sensitive entities — SSNs, cards, keys, emails, phones, and custom patterns.
- 02
Redact or block
Matched spans become high-entropy placeholder tokens; rules can block the prompt entirely. The provider never sees raw values.
- 03
Govern & route
The allow-list and org context are applied server-side, and the resolved model is recorded — enforced, not merely suggested.
- 04
Restore & record
Tokens are restored to real values in your view; one immutable audit record captures exactly what happened.
Capabilities
Everything a regulated team needs in v1
Streaming governed chat
A fast, familiar chat experience with saved conversations and resume — token-by-token streaming so it feels as good as the tools your team already reaches for.
Inline redaction & blocking
Detectors scan each prompt for PII, secrets, and custom patterns. Matches are redacted to placeholder tokens, or the prompt is blocked outright — before egress.
File upload, scanned
Attach PDFs, DOCX, TXT, or CSV. Their contents flow through the same redaction pipeline before a single byte reaches a provider.
Model allow-list
Admins choose exactly which providers and models the org may use, and whether end users can pick. Disallowed models are rejected server-side.
Prepend org context
Apply standing instructions and guardrails to every outbound prompt as a system-level message — versioned, and referenced by each audit record.
Per-prompt audit & metering
One immutable record per prompt — requested vs. resolved model, redactions, tokens, and estimated cost — feeding usage reporting by user and model.
Redaction your team barely notices
Most DLP makes the AI feel broken. Wardary uses reversible tokenization: sensitive spans are swapped for stable placeholders before the call, and restored to their real values when the reply comes back — so the conversation reads seamlessly while the provider only ever saw tokens.
- Placeholders are high-entropy per-request nonces
- The token↔value map lives in memory only and is never persisted under a no-retain rule
- Restore refuses unknown tokens, so a hallucinated placeholder can never inject a real value
Why this matters
A resumed chat shows the masked placeholder for any span your policy says must never be stored — while the live session showed the real value. Persistence follows your retention rules, span by span, with most-restrictive-wins on overlaps.
Under the hood
Built so the guarantee is easy to hold
- Single egress seam
One way out
Every provider call goes through a single callProvider() seam, with a test asserting nothing else imports a provider SDK. Redaction always runs first.
- In-process pipeline
No extra hops
Redaction and audit stay in-process where the leak-free guarantee is easiest to enforce — not bolted on as a downstream service.
- DB-level immutability
Append-only, for real
Audit immutability is enforced at Postgres, not in middleware — so “append-only” survives an auditor’s scrutiny.
Roadmap
What ships now, and what's next
- Governed streaming chat + saved conversations
- Inline redaction & blocking (prompts + files)
- Model allow-list & prepend context
- Per-prompt audit + usage metering
- Human redaction review & override queue
- Contextual rule-based routing engine
- Local / self-hosted model targets
- Browser/endpoint guard for direct AI use
We sequence honestly and gate new surfaces on real customer pull.