KEEL Core

Kernel for Efficient Execution & Latency

KEEL Core is a sub-10ms memory-safe Rust proxy that treats prompt evaluation as an in-memory classification problem solved before inference starts.

02 · Classifier

Classifier

In-process INT8 ONNX intent classification in under 8ms. Reasoning depth, tool-use likelihood, and safety class are scored before a single token is billed.

p99 classification: 7.8ms

Rust Runtime

Tokio Async Engine

Multi-threaded work-stealing scheduler with lock-free ring buffers.

Classifier

ONNX INT8 Model (22MB)

In-process execution with zero IPC network hops. P99 classification < 8ms.

Latency Budget

< 10ms Total Overhead

Hard-bounded execution budget before routing decisions commit.

Context Engine

SIMD-Accelerated Pruning

Content-hashes system prompts to eliminate redundant middle-turn tokens.

Telemetry Plane

Out-of-Band ClickHouse Pipeline

Asynchronous streaming logs that never block the client TTFT stream.

Deployment Matrix

Single Binary / Helm VPC

Deployable as a stateless Docker container across EKS, GKE, or AKS.

Deep Dive

KEEL Core Low-Level Engineering Primitives

What the proxy actually does between accepting your socket and committing a routing decision.

Lock-Free Concurrency & Memory Safety

Built using Rust's async task model on Tokio. Memory allocations for request buffers are pre-allocated at startup. Zero heap re-allocation during hot routing execution paths.

In-Process Machine Learning Inference

Uses direct C++ bindings to the embedded ONNX Runtime. The 22MB INT8 quantized intent classifier executes directly in CPU memory cache without calling external API endpoints.

Deterministic Circuit Breaker & Fallbacks

Executes a lightweight streaming JSON schema parser. If an SLM outputs malformed tokens or exceeds TTFT thresholds, the connection transparently escalates to GPT-4o.

Inspector

Walk the request lifecycle, step by step

Select a stage to inspect simulated latency, token reduction, and runtime memory behavior.

INT8 intent classification in CPU cache

The 22MB quantized model scores reasoning depth, tool-use likelihood, and safety class in-process — no IPC, no network hop.

P99 classification

7.8ms

Model size

22MB INT8

Tokens billed

0

Stop Paying Top-Tier Prices for Basic String Utilities.

Route intelligently, prune aggressively, and reserve frontier reasoning for the requests that actually need it. Two lines of code, sub-10ms overhead.