AI stopped answering. It started acting.
Assistants file tickets, move money, and message citizens. Agents call tools on their own. Every one of those is a decision your organization owns, whether you saw it or not.
AI governance & assurance
Your AI makes decisions.
Turrigan makes them defensible.
One inline chokepoint governs every AI output and every agent action: allow, block, or escalate, in milliseconds, with a tamper-evident record an auditor can verify years later. Not a dashboard that watches. A sentry that acts, and proves it.
Deterministic core. Runs air-gapped with no external AI service. Milliseconds, not seconds, because there is no LLM in the hot path.
// the AI was about to output:
"Your card 4111 1111 1111 1111 is on file."
// POST /v1/govern responded:
{
"verdict": "block",
"delivered": false,
"triggered_controls": ["eu-ai-act:art-10", "eu-ai-act:art-12"],
"detectors": [{ "detector": "pii", "flagged": true }]
}
// the stored record is redacted + hash-chained:
"output_redacted": "Your card [REDACTED:CREDIT_CARD] is on file."The uncomfortable part of shipping AI
The regulators arrived first.
The EU AI Act, NIST AI RMF, and ISO/IEC 42001 all ask the same question: show us how each decision was controlled. "We have a policy PDF" is not an answer. Logs you can edit are not evidence.
One bad output is a headline.
A leaked ID number, a fabricated legal citation, a prompt-injected agent with payment access. The cost of the thousandth output failing is not covered by the first 999 going fine.
How it works
Three steps. Milliseconds. Proof forever.
1
Submit the decision
Before your AI's output reaches a user, or before your agent executes a tool call, send it to /v1/govern. One HTTP call with a scoped, revocable key.
2
The floor decides
A deterministic detector floor screens for leaked personal data, fabricated citations, prompt injection, unsafe content, excessive agency, and coverage gaps. No cloud LLM in the hot path. Critical findings block; uncertainty escalates to a human. It fails closed, never open.
3
The evidence writes itself
Every decision lands in an append-only, hash-chained log: redacted, encrypted, explainable, and mapped to the exact controls it satisfies. Tampering is mathematically detectable. Erasure is crypto-shredding that leaves the chain verifiable.
One platform, three planes
Govern what AI says, prove it to auditors,
and control what agents do
Plane A
Content governance
Eight detectors, English and Arabic, screen every output inline: personal data (with redaction spans), verifiable-claim checking against a curated offline legal registry, prompt injection, safety, uncertainty abstention, and fairness tagging.
- Deterministic, offline, ReDoS-safe
- Hardened against five 1,000-case adversarial corpora across the floor
- No LLM in the hot path, so verdicts land in milliseconds
Plane B
Compliance & assurance
The differentiator. Decisions are not just logged, they are mapped to controls in six versioned frameworks. Live conformity posture, per-control evidence, one-click Annex IV core export, drift gates, and Article 73 incident workflows.
- EU AI Act, NIST AI RMF, ISO/IEC 42001
- OWASP LLM Top 10, GDPR, MITRE ATLAS
- Right-to-explanation on every decision
Plane C
Agentic action governance
Agents submit proposed tool calls before executing them. Tool allow-lists block unknown tools, privilege-scope checks catch overreach, parameter-risk screening flags dangerous arguments. Turrigan authorizes the decision to act. It never executes anything itself.
- Addresses OWASP LLM06: Excessive Agency
- Same chain, same evidence, same oversight
Evidence, not assertions
The audit that takes weeks becomes an export that takes seconds
Ask any AI vendor how they will support your next audit. Then ask them to prove a log entry was never edited. Turrigan's decision log is hash-chained and signed into verifiable checkpoints: any edit, deletion, or reorder is detected, with the exact break position.
- Live conformity posture. Covered and not-covered per control, honestly reported, never inflated.
- Annex IV core pack. The EU AI Act technical-documentation skeleton, generated from real decisions.
- Per-decision explanation. A deterministic, human-readable reason for every verdict. Article 13 and 86 ready.
- Data minimisation built in. Content is redacted before storage and sealed with per-record keys. Raw personal data never appears on any read surface.
Sovereignty is a feature
Runs where your data is allowed to live
Managed cloud
Fastest start. Your tenant, your keys, your encrypted data, our operations. Physical database isolation available per tenant.
Your database
Bring your own datastore: the decision log lives in a database you host and own. The connection secret is sealed at rest. Walk away with your data any day.
Air-gapped / sovereign
The deterministic floor needs no external AI service at all. Government and regulated deployments run fully on-premises, offline, with customer-managed keys and crypto-shredding.
Read this before any vendor demo
What we deliberately do not claim
We do not make you compliant.
No tool can. Compliance is an organizational outcome. Turrigan produces the per-decision evidence and controls infrastructure that makes your compliance case provable instead of aspirational.
We do not certify fairness.
Our disparate-impact screening surfaces signals worth investigating, population-level and honest about its limits. Anyone selling you push-button fairness certification is selling you risk.
This honesty is not modesty. It is the product. An assurance layer that inflates its own coverage is worse than none, so every posture Turrigan reports is one an independent auditor can reproduce. The evidence infrastructure has passed an independent technical review against the EU AI Act, alongside adversarial reviews across the detector floor. Formal notified-body certification remains a separate, later step, and we will never pretend otherwise.
For your engineers
One endpoint to adopt.
Not a platform migration.
Keep your models, your vendors, your stack. Governance is one HTTP call in the path you already control, with typed schemas for every response in the interactive API reference.
- Scoped, expiring, revocable API keys
- Fails closed: an unreachable datastore can never silently allow
- Per-tenant rate isolation, 200 KB bounded inputs, no LLM in the hot path
curl https://api.turrigan.com/v1/govern \
-H "X-API-Key: trg_..." \
-H "Content-Type: application/json" \
-d '{
"input": "user question",
"output": "candidate AI answer",
"profile": "customer-support"
}'
# → {"verdict": "allow", "delivered": true, ...}Pilot Turrigan on your real AI traffic
We are onboarding a small number of design partners: a free, time-boxed pilot on a dedicated, isolated tenant, with a scoped key and your data in a database that can be handed to you or cryptographically shredded when it ends. You keep the evidence either way.
Tell us what your AI does and what a bad day looks like. We will tell you honestly whether Turrigan helps. Pilot scope and terms are confirmed in writing before onboarding; nothing here is a binding offer.