Technical reference / 04
Local inference, auditable orchestration, and institution-controlled integrations — designed to keep security reasoning inside your network boundary.
System architecture
AegisGrid is a three-layer system. The Local Inference Cluster runs the LLMs and vector store that produce all reasoning; the Agent Orchestrator sequences the multi-agent workflow and emits the audit log; the Integration Layer talks to the systems you already operate (ServiceNow, SIEM, email). The three layers share a host, a network, and a perimeter. There is no off-host component of the AegisGrid runtime itself. There is no third-party API call in the runtime path.
The architecture is designed to be auditable. Every agent tool call, prompt, and decision is written to an append-only, hash-chained audit log. Every log entry is signed. Every entry is replayable. The containment switch is a millisecond-time primitive, not a request ticket. The full data flow — what enters, what is processed, what leaves (nothing) — is documented below and cross-referenced to the compliance crosswalk on /compliance.
Deployment topologies
No network path between AegisGrid and the outside world. All updates are delivered as signed offline packages.
Inference, audit log, and integrations all operate inside the air gap. This is the topology we recommend for institutions operating under the most restrictive data-handling regimes.
network: isolated
updates: signed offline package
runtime: local inference
audit: on-host / immutable
AegisGrid runs inside a dedicated VPC on Azure GovCloud, AWS GovCloud, or an equivalent government-grade cloud.
Network egress is restricted to the institutional SIEM, ServiceNow, and the signed-package update endpoint. This is the topology most institutions select first.
network: dedicated VPC
egress: allowlisted connectors
compute: GPU-backed nodes
updates: signed package
AegisGrid runs on hardware you own and rack. We provide a reference hardware specification (GPU class, RAM, storage) and a deployment runbook.
The on-prem topology is the most common choice for R1 research universities with their own data centers.
network: institution-owned
compute: GPU class reference
storage: encrypted local
deployment: runbook + handoff
Private cloud and on-prem are deployment patterns established in NetWit's platform language; AegisGrid applies that pattern to a local-inference security runtime. Source: NetWit platform
The agentic stack
AegisGrid is built on standard open-source components. None of these are certifications — they are the technologies the runtime is composed of.
Security model
agent→agent and agent→tool traffic uses mutual authentication and encryption in transit.
each entry is signed and hash-chained to the prior entry, making tampering detectable.
role scope controls which workflows, audit entries, and overrides each human user can access.
a designated role can revoke agent tool authority and rotate runtime credentials; the action is logged.
Data flow
Documents are parsed, controls are extracted, policy mappings are produced, risk ratings are drafted. All processing happens on the AegisGrid host. Work notes, queue updates, and audit entries can flow back into ServiceNow and the SIEM; no document content, extracted control, or risk-matrix output is sent to any third-party API.
Next: inspect the evidence
The proof workflow shows the artifacts, lineage, and audit entries an institutional team can review before deployment.