Policy at runtime. Under 75 ms. p99.
Observation is not control. The ShadowIQ Gateway sits inline on every model call — blocks prompt injection, redacts PII, applies your policy-as-code, and calls a human only when the edge case demands it.
Inline. Fast. Deterministic. Every call, every model, every tenant.
Three moves, fully automated.
No long onboarding, no hand-rolled detection rules. ShadowIQ ships with defaults tuned to the regulatory floor — customize only where your risk appetite demands.
Point traffic at the gateway.
One endpoint for OpenAI-compatible, Anthropic, Bedrock, Vertex, Azure OpenAI, or bring-your-own. SDKs for TypeScript, Python, Go — or drop-in at your existing proxy.
Declare policy-as-code.
YAML or Rego. Version-controlled, tested, and reviewed like any other deployable artifact. Roll out by tenant, environment, or workload.
Run, measure, tune.
Live latency budget, per-policy hit rate, false-positive tracker, and a shadow mode that observes before it enforces.
Every control a regulator or auditor will ask about.
Sub-75 ms p99
Parallel evaluation, pre-compiled policy WASM, and warm tenant pools keep the 99th percentile tight even during burst traffic.
Prompt injection defense
Multi-classifier ensemble (rule + small model + LLM-judge quorum) tuned on 2,400+ labeled adversarial samples.
Redaction & tokenization
SSN, passport, PAN, and customer-schema detectors. Redact, hash, or tokenize inline — decision recorded.
Allowlist + deny
Model-of-record policies per tenant, per environment. Block OpenAI in Germany, Azure EU-only in Frankfurt — all config.
Human-in-the-loop
Risk-tier a decision to a human reviewer. Approvals flow through Slack or ServiceNow and close the policy loop in seconds.
Observe before enforce
Shadow mode mirrors every call, logs would-be decisions, and tunes false-positive rate without impacting users.
Output filtering at stream
Token-level output filters. Stops an unsafe answer mid-stream without tearing the entire response apart.
SaaS · VPC · self-host
Same binary. Same policy bundle. Same evidence format. Pick the deployment that passes your procurement review.
Multi-tenant residency
Hard tenant isolation, regional pinning, FIPS-validated crypto, and per-tenant HSM keys available.
Answered by the architecture, not the sales deck.
Configurable fail mode per workload — fail-closed (block) for regulated workloads, fail-open with alert for less sensitive ones. Availability target: 99.99% with regional failover.
Yes. Self-hosted deployment ships as a signed container bundle with its own key material. Evidence anchoring uses an internal transparency log rather than our public service.
Policy-as-code tests run in CI like any other code. You get a signed test report plus shadow-mode telemetry from staging workloads before promotion to production.
No. Customer prompts and responses never enter our training data. This is contractual (DPA) and architectural (isolated evaluation environment with no egress to training infrastructure).