Architecture
An architecture review board will not accept “trust us” for anything running inside your environment. This is what deploys, what licenses it, what data it touches, and what breaks when it stops.
01Distribution
“One thing to install” describes the experience, not who owns a copy of the software. AI Pulse does not embed, fork or ship copies of any third-party component.
The deployment chart references the official upstream images by pinned tag or digest. At install time your own infrastructure — or SYLVA, running that same chart on your behalf — pulls those images directly from the vendor's own registry. For an air-gapped estate, your IT mirrors the pinned images into your internal registry first and the chart points there instead.
That is the difference between redistributing third-party software, which raises real licensing questions, and configuring open source you obtained yourself, which mostly does not. What SYLVA ships is its own components and its own authored configuration.
On the open-core question. Two of the components are open-core projects: a permissive base with a separate commercial tier of that project gating its own SSO, multi-user administration and audit retention. Everything AI Pulse needs is in the free, MIT-licensed core. Those paid tiers are never invoked, because no person ever logs into either tool's own console — only AI Pulse's backend talks to them, and every human uses AI Pulse's own interface and access control.
02Component by component
| Component | What it is for | Licence | Data it touches | If it fails |
|---|---|---|---|---|
| AI GatewayOpen source · LiteLLM | The mandatory path to every model call. Chosen for the widest provider coverage in one place, including self-hosted OpenAI-compatible endpoints. | MIT | Prompts and completions, in transit | High |
| Trace StoreOpen source · Langfuse | Durable, searchable record of every interaction. The most sensitive store in the deployment, and subject to the retention policy. | MIT | Prompts and completions, at rest | Medium |
| Safety ScannerOpen source · Presidio | Deterministic detection of personal data and secrets on the hot path. Pattern-based rather than model-based, so it is fast enough to sit inline. | MIT | Content in memory, at the moment of the check | Medium |
| Trace CollectorOpen source · OpenTelemetry | Receives the deeper traces from agents that opt into them. Entirely optional enrichment. | Apache-2.0 | Framework-native trace data | Low |
| MCP BrokerSYLVA build | In-path control on MCP execution: authenticates the agent, attests the live server against an approved baseline, scans arguments, records the decision. Fails closed. | Proprietary | Tool arguments, in transit | High |
| Action GatekeeperSYLVA build | Inline on the sensitive-action path. Permits, denies or pauses for approval; it never holds a credential and never executes anything. | Proprietary | Action parameters — never a credential | High |
| Policy EngineSYLVA build | Evaluates declared rules as data rather than code, so rules can be extended without a build. Degrades to “not assessed”, never to a false pass. | Proprietary | Configuration and attestation state | Low |
| Remediation BridgeSYLVA build | Routes findings to self-heal, ITSM, SIEM or a signed webhook. Holds no standing authority beyond a pre-approved list. | Proprietary | Finding context | Low |
| AI Pulse applicationSYLVA build | The registry, governance ledger, attestations, framework mapping, evidence vault and every dashboard. | Proprietary | Registry, ledger, evidence | Medium |
Four are open source under permissive licences at no cost. Five are SYLVA's own build — and they are the five that do the governing.
03What we do not deploy
LangChain, LangGraph, LlamaIndex, Semantic Kernel, CrewAI, AutoGen and Pydantic AI are whatever your teams already chose and licensed. AI Pulse installs none of them. The only relationship is passive: if an agent points its own tracing at our collector, we can read what it emits.
ServiceNow, Jira Service Management, Splunk, Microsoft Sentinel — integrated with, never installed. The connectors talk to your own licensed instances.
No agent runs on a server anywhere in your estate. On Azure the footprint is a managed application and a read-only identity; the enforcement services run as containers you deploy.
The full justification — every component's alternatives considered, permissions and failure behaviour — is shared under NDA during evaluation. Request the pack →
