Build compliant AI in minutes.
Everything you need to integrate ZeroDrift: quickstart guides, full API reference, SDK documentation, and a library of pre-built compliance policies.
Where do you want to start?
Quickstart
Intercept your first AI response in under five minutes. Install the SDK, set an API key, and you are live.
API Reference
Complete reference for every endpoint: intercept, policies, audit log, and webhook configuration.
SDK Guides
Drop-in SDKs for Python, Node.js, and Go. Copy a working example and customize from there.
Policy Library
Pre-built policy sets for financial services, healthcare, and regulated data. Activate with a single API call.
Key concepts
Understanding how ZeroDrift works helps you build more effective compliance configurations.
The intercept loop
Every AI response passes through ZeroDrift before reaching your user. The SDK proxies the model call, sends the raw output to our guardrail layer, and returns either the original or a policy-compliant rewrite, all within a single await.
Policies and rule sets
A policy is a named collection of rules that defines what is and is not allowed in a response. Rules match on topic, phrasing pattern, regulatory category, or a combination. You can activate multiple policies at once and they evaluate in priority order.
Rewrite vs. block
When a rule triggers, ZeroDrift either rewrites the violating passage into compliant language or blocks the full response and returns a safe fallback. You configure the action per rule. Rewrites preserve the useful parts of the response; blocks are for zero-tolerance cases.
Audit log
Every intercept event is recorded: what the raw model said, which rule triggered, what action was taken, and the final response delivered. The audit log is queryable by time range, policy, rule, user session, and action type, and exportable to your SIEM or compliance toolchain.
Context passthrough
You can attach arbitrary metadata to each intercept call, such as user role, jurisdiction, or session type. Rules can reference context fields, so the same policy can behave differently for an authenticated advisor versus an anonymous user on the same platform.
Latency and availability
The ZeroDrift guardrail layer adds under 30 milliseconds at p99 in our US-East region. Calls are processed asynchronously in a fire-and-verify pattern so your model response stream is never blocked waiting for an evaluation result.