How Arkheia deploys

§ 01 — Deployment options

Choose where you need control. Each deployment mode has a defined trust boundary and information flow. Expand any card to see the architecture diagram with animated flows.

§ 02 — Enterprise vLLM integration

Zero-modification deployment

Arkheia deploys as a transparent sidecar proxy between your application and your existing vLLM inference server. No code changes, no SDK, no model modifications. Your application talks to Arkheia exactly as it would talk to vLLM directly — the OpenAI-compatible API contract is preserved.

# Before
OPENAI_BASE_URL=http://vllm:8000

# After
OPENAI_BASE_URL=http://arkheia:8080

{ "arkheia": { "risk": "LOW", "confidence": 0.92, "strategy": "ensemble" } }

§ 03 — System topology

Five microservices

Five FastAPI microservices communicate over internal HTTP. All services are independently deployable and fail gracefully — if any support service goes down, detection continues.

§ 04 — Telemetry flow

Signal without content

The on-prem engine derives a runtime signal from two observable surfaces. Telemetry is collected locally, processed within the detection engine, and emitted as a risk signal. No prompts, responses, or application content are inspected or retained.

§ 05 — Profile system & adaptive routing

Zero code changes for new models

Model profiles are YAML files loaded at startup and hot-reloaded at runtime. The adaptive router selects the optimal profile variant based on actual response characteristics.

§ 06 — Operational tenets

Architectural, not configurable

§ 07 — Security architecture

Defence-in-depth

API key authentication, license verification, and hashed credential storage. Designed for air-gapped enterprise deployments.

§ 08 — Technology stack

Full stack reference