Deployment

Deploy EL-Ai-NE where privacy, latency, continuity, and operational control require it.

Start with a proof of concept, move to a pilot at a real site, then roll out by cohort. Every track and every deployment profile preserves the same identity, policy, workflow, action, verification, and evidence model.

Rollout tracks

Proof of concept, pilot, then production.

The architecture is the same on every track. What changes is the scope, the data, the resilience and the standard of evidence. Passing a gate at a small scope never qualifies a larger one.

Track 1

Proof of concept

Thirty days in a reference lab on synthetic or expressly permitted data.

You need
One virtualization host that runs five required and two optional virtual machines, with an emulated branch, headquarters and WAN.
You prove
A governed answer from a local model, the refusals, the emergency stop, a deletion that survives a restart, and one bounded, approved action.
You leave when
Nine pass-or-fail gates are met and a rollout decision is accepted.
Track 2

Pilot

The same loop on real sources at a representative site, then at 10 to 25 diverse sites.

You need
One site node beside your existing switch and WAN edge, and a small control deployment with a tested backup and restore.
You prove
Value on real work, read-only first, with source rights recorded for every source.
You leave when
The exit criteria for the cohort are met: security, quality, completion, verification, capacity, cost, outcome and recovery.
Track 3

Production

Rollout by cohort: 50 to 100 sites, a full region, several hundred sites, then the estate.

You need
A control and evidence cell across separate failure domains, regional cells, and a hardware class sized to each site.
You prove
Each capability for the exact release, models, hardware and topology you will run.
You keep
Separate qualified and enabled records, runbooks, tested recovery and periodic requalification.

What changes between the tracks

Design choiceProof of conceptPilotProduction
DataSynthetic or expressly permitted.Real sources, read-only first, each with a recorded rights grant.Classified, retained and revoked under accepted policy.
High availabilityNone.Not required. Recovery is defined and tested.Required, across named failure domains, with tested objectives.
IdentityLocal accounts.Local identity with hardened accounts.Your enterprise identity provider.
How sites connectApplications call the gateway, or use it as an explicit proxy.The same, with selected service chaining after path and bypass tests.The accepted insertion profile for each site.
ActionsOne bounded, approved action.One bounded, approved action class.Qualified action classes under an owner-approved consequence ceiling.
External AIOff.Off.Off unless a destination, purpose and data class are admitted.
Evidence standardRehearsal.Rehearsal, then cohort exit criteria.Counted qualification.

The nine gates that end a proof of concept

  1. No unauthorized accessNo unauthorized tenant or site access in negative tests.
  2. Every change is accountableEvery change has an identity-bound request, a policy decision and a receipt.
  3. No external bypassNo external route bypasses data-loss controls and provider policy.
  4. Every success is verifiedEvery successful action has an independent verification.
  5. Unknown stays unknownMissing critical evidence produces unknown or degraded, not success.
  6. Rollback is realRollback or containment has been exercised.
  7. The measure is agreedThe business baseline and outcome method are accepted.
  8. Cost is honestCost is measured, or clearly marked as predicted.
  9. Operators stay in controlOperators can see and control the workflow.

What a proof of concept can show today. Sign-in, chat on a local model, the policy-first gateway, the operational record, the governance ledger, tombstones, the emergency stop and usage metering are Available now. Approved actions are Included, enabled per qualification and need a signed action allowlist. High availability, cross-site state and disconnected operation are Planned. See each use case for its own status.

Without and with EL-Ai-NE

See how placement changes the operating path—not the authority model.

Use the animated network lab to compare direct requests, background workflows, endpoint control, SD-WAN insertion, disconnected operation, and cross-site content reuse.

Without EL-Ai-NEPoint tools and manual handoffs
User or eventSeparate toolsManual actionPartial evidence
With EL-Ai-NEGoverned, local-first, verified operation
Request or eventIdentity + PolicyEL-Flow + ActionVerified outcome

Interactive deployment profiles

Placement changes. Authority does not.

Select a profile to see the high-level path. The design and implementation guides define interfaces, addressing, DNS, DHCP, NTP, certificates, firewall rules, MTU, routing, failure behavior, and acceptance testing. Start with applications calling the gateway directly or using it as an explicit proxy; add service chaining for selected traffic only after path and bypass tests.

Network prerequisites

Production readiness begins with ordinary infrastructure done well.

EL-Ai-NE depends on reliable identity, name resolution, time, certificates, addressing, routing, storage, observability, and rollback. These are implementation requirements—not hidden assumptions.

IP addressing and interfaces

Static or reserved management addressing for infrastructure; documented user, management, regional, and service-transit networks; explicit ownership of the service-in and service-out interfaces where used.

DNS and certificates

Internal service names, forward and reverse records, certificate SAN alignment, split-horizon behavior, revocation, and secure endpoint enrollment.

DHCP and NTP

Gateway, DNS, domain/search list, and NTP options where applicable. Accurate time protects certificates, action leases, event ordering, report windows, and evidence.

Routing, firewall, and MTU

Outbound TLS, telemetry, artifact, and model routes; symmetric service paths; proxy and NAT policy; PMTUD; QoS; bypass or fail behavior; and failure-domain testing.

Proxy, cache, and content placement

Define explicit, transparent, reverse-proxy, or connector participation; cacheable traffic classes; TLS policy; local and regional storage; trust groups; hotset quotas; freshness; and tombstone behavior.

What each site needs

The lowest hardware class that fits the work.

EL-Ai-NE defines capability classes, not vendor products. A site receives the lowest class that satisfies its measured workload. Site count alone does not decide where accelerators go.

ClassUse it when the siteStarting specification
Core siteNeeds policy, cache, connectors, evidence and light classification, with larger answers served from the region.8 to 16 cores, 32 to 64 GB memory, 2 TB fast storage. No accelerator required.
AI siteNeeds local extraction, retrieval, agent support or selected vision, or must keep inference on site.12 to 24 cores, 64 to 128 GB memory, 4 TB fast storage, one 24 to 32 GB accelerator where evaluation justifies it.
AI Plus siteHas high volume, multimodal work, privacy isolation, manufacturing or complex local reasoning.16 to 32 cores, 128 to 256 GB memory, 4 to 8 TB fast storage, a 48 to 96 GB accelerator class.
Critical siteIs high-consequence or must run through prolonged isolation.A high-availability pair with redundant power, network and storage.
Regional AI cellServes pooled larger-model inference, overflow and site failover for a group of sites.A multi-accelerator cluster with one node of spare capacity.
Control and evidence cellHolds the Manager, policy, the operational record, evidence, cost and observability for the estate.Processor, memory and storage focused, placed across separate failure domains.

Starting points from the baseline planning envelopes, not certified bills of materials. Optional cells for regional storage, training and isolated browser work are added only when a profile needs them. The Bill of Materials lists minimum and recommended specifications with example hardware for every class.

Around every site node

  • Three connections to startLights-out management, a management port and a service port on your existing switch.
  • Two more for service chainingA service-in and a service-out port on the WAN edge, used only for selected traffic.
  • Protected powerA UPS sized to the node, with separate circuits for dual power supplies.
  • Trusted hardwareA security chip, verified boot, encrypted storage and a locked enclosure.

Baseline deployment reference

A concrete lab plan with explicit customer inputs.

The 10.70.0.0/24 values below are isolated lab examples. Manager and Edge VMs can run outside EVE-NG, with lab clouds connecting the intended service paths. Replace addresses and map actual MACs, vNICs, VLANs, routes, DNS, DHCP reservations, NTP and certificate names before use.

RoleReference valuePurpose
Manager10.70.0.10TCP 443 management ingress
Regional inference10.70.0.20Qualified internal route
Branch / HQ / DC Edge10.70.0.31 / .32 / .33Separate management identities
Gateway / DNS / NTP10.70.0.1 / .2 / .3Accepted management services
Branch and HQ service portsGi1/2 IN / Gi1/3 OUTExplicit service-in and service-out roles
DC service portsGi3/0 IN / Gi3/1 OUTKeep management separate

Reference lab sizing

Virtual machineAllocationPurpose
Manager8 vCPU, 16 GiB, 120 GiBControl services and console
Data8 vCPU, 32 GiB, 500 GiBDatabases, evidence and event storage
Regional8 vCPU, 24 GiB, 250 GiBModel host and regional services
Two site nodes4 vCPU, 16 GiB, 150 GiB eachBranch and headquarters sites
Test services2 vCPU, 8 GiB, 100 GiBSynthetic sources and a test ticketing system
Isolated workspace host (optional)8 vCPU, 24 GiB, 200 GiBWorkspace isolation tests

The lab allocates 42 vCPU, 136 GiB of memory and 1,470 GiB of disk in total. It is a starting allocation for isolated testing, not a performance guarantee, and it does not establish high availability.

Begin with selected internet traffic, preserve SD-WAN control and approved DIA breakout, and verify return-path symmetry, MTU, bypass and rollback. Do not bridge service interfaces through the management switch.

Request the Architecture Guide

Deployment scope

Choose business scope and physical placement separately.

Business profiles progress from a governed pilot through regional operations, controlled actions, distributed enterprise operations and controlled autonomy. Placement profiles are lab, site, regional, distributed, air-gapped and assurance. A qualified small profile does not imply qualification for the full 1,000-site design target.

Capacity, latency, local completion, availability and recovery targets are accepted and measured for the named workload. Optional features remain disabled until their dependencies and qualification evidence are complete.

Ask EL-Ai-NEPublic website guide
Ask about capabilities, governance, privacy, Brain Trust, risk, deployment, pricing, comparisons, or public resources.