Arkheia

On-Premise Detection Engine

Deploy inference-time risk monitoring alongside self-hosted LLMs. Kubernetes-native. Fail-open by default.

Request Demo →View API Proxy

A Detection Layer for Self-Hosted LLMs

The Arkheia On-Premise Engine runs inside your environment alongside your inference stack. It observes model behavior during generation and emits an enriched response that includes a risk signal.

Arkheia does not judge correctness or validate content. It surfaces behavioral signals that allow systems and operators to decide when additional scrutiny is warranted.

The constraint

Self-hosted LLM deployments offer limited visibility into inference-time behavior. Outputs that appear confident may still be produced under internal uncertainty.

Why it matters

In regulated environments, the lack of runtime observability creates operational risk and can block AI adoption entirely.

// Enriched response

{
  "choices": [...],
  "arkheia": {
    "risk_level": "MEDIUM",
    "confidence": 0.62,
    "flags": ["signal_anomaly"],
    "recommendation": "review_or_route"
  }
}

§ Deployment Architecture

Alongside your stack — not embedded in it

Arkheia deploys as a transparent sidecar proxy. Your application and vLLM remain entirely unchanged. Traffic routes through the sidecar during inference, signals are extracted and discarded, and the enriched response returns with the risk score appended.

ENTERPRISE INFRASTRUCTURE BOUNDARYYour AppAPI: unchangedARKHEIA SIDECARTransparent proxySignals extracted — never storedfail-open · content never accessedprofile version logged per callvLLM / Modelunchangedrequestforwardsenriched response: standard output + risk score

If the sidecar is unavailable, traffic passes through directly — fail-open by design. No single point of failure is introduced into your inference path.

Three Steps to Deployment

1

Deploy the Sidecar

Install Arkheia alongside your inference stack via Helm chart or Kubernetes manifests.

helm install arkheia arkheia/detection-engine
2

Point Inference Through Arkheia

Route requests through the engine inside your cluster. Change one line in your config. No SDK, no model retraining.

OPENAI_BASE_URL=http://arkheia:8080
3

Receive Enriched Responses

Get standard responses plus risk metadata and operational recommendations. The OpenAI-compatible API contract is preserved.

Integration time is typically one to two days, depending on your inference stack and governance requirements.

Designed for Enterprise Workflows

Zero Data Exfiltration

Processing occurs entirely inside your infrastructure. Prompts and responses are strictly processed within your environment and never persisted out-of-band.

Fail-Open by Default

If Arkheia is unavailable, inference proceeds normally. No single point of failure in your critical path.

Low Overhead

Designed to add minimal latency to inference pipelines. Detection runs alongside response delivery.

Hot-Swappable Profiles

Update detection profiles without service interruption. Arkheia monitors model drift over time — when a model's behavioural baseline shifts, recalibrated profiles are deployed live without restarts.

Model-Independent

Works with vLLM-compatible deployments and common inference stacks. No model weight changes required.

Audit Trail

Per-invocation audit log: timestamp, model ID, active profile version, risk decision, and confidence. Full traceability from each API call back to the YAML profile and characterisation corpus.

Validated Where It Matters

On-prem deployments avoid dependency on commercial API telemetry and can be governed entirely within your environment.

ON-PREM

vLLM

Supported inference stacks

MODEL FAMILY

Llama / Mistral / Qwen

Common self-hosted targets

POSTURE

Fail-Open

Designed for critical paths

Deploy Inference-Time Observability

Arkheia is designed to sit quietly alongside production inference — surfacing signals without disrupting workflows.

Talk to Arkheia →Security & Privacy