Loading...
Loading...
The Engine
Akretic acts as a transparent reverse-proxy. We deserialize the LLM's network stream, parse the JSON functionCall intent, and evaluate the Cedar AST in memory before the payload ever reaches your internal APIs. Zero application rewrites. Zero SDK imports.
Memory Safety
Compile-Time Enforced
Intercept Latency
< 1ms
use axum::{extract::Json, http::StatusCode, response::IntoResponse};
use serde::Deserialize;
use serde_json::json;
#[derive(Deserialize)]
struct ToolCallPayload {
function: FunctionCall,
model: String,
}
#[derive(Deserialize)]
struct FunctionCall {
name: String,
arguments: serde_json::Value,
}
/// L7 intercept handler. Evaluates Cedar AST before proxying.
pub async fn intercept_tool_call(
Json(payload): Json<ToolCallPayload>,
) -> impl IntoResponse {
// 1. Parse the functionCall intent from the LLM stream.
let intent = &payload.function;
// 2. RPC to Gate0 for deterministic Cedar AST evaluation.
match gate0_client::evaluate_cedar(
&intent.name,
&intent.arguments,
).await {
PolicyResult::Allow => {
// 3. Forward to upstream API.
proxy::forward(payload).await
}
PolicyResult::Deny(reason) => {
// 4. Log denial to Iron Ledger, return 403.
ledger_client::record_denial(&reason).await;
(StatusCode::FORBIDDEN, Json(json!({
"error": { "message": reason }
}))).into_response()
}
}
}Zero-allocation, panic-free micro-policy engines. Rules are evaluated in a fixed sequence with a strict deny-overrides strategy. Transaction verification is predictable and sub-millisecond, eliminating allocator jitter in the policy evaluation path.
Policy Gate
Deterministic Enforcement
Cedar AST evaluated before upstream delivery
Evaluation
<0.3ms
Sub-millisecond deterministic verification
Strategy
Deny-Override
Fixed-sequence rule evaluation with strict denials
Architected in memory-safe Rust to deploy seamlessly across heterogeneous environments—from Oracle Cloud (OKE) and AWS (EKS), to massive NVIDIA GPU clusters. The Akretic control plane enforces deterministic security without introducing latency to your underlying inference engine, natively interoperating with architectures like NVIDIA NIM™ and Oracle OKE.
Deployment Targets
4+
AWS EKS, Oracle OKE, Azure AKS, & Bare-Metal
Proxy Overhead
<1ms
Zero-bottleneck deterministic evaluation
LLM Compatibility
Model Independent
Secures open-weight and proprietary models
*Metrics reflect benchmarked Gate0 evaluation latency under maximum parallel throughput.