System resource
Rolling 3σ baselines over CPU and memory, computed separately for in-cluster workloads and remote hosts, so “normal” is learned per machine rather than assumed.
Autonomous site reliability engineering
Infrastructure that diagnoses itself — and proves what it did.
Omni is a multi-agent SRE platform for Kubernetes and Linux fleets. It collects evidence through four independent diagnostic lanes, reasons over it with a locally hosted LLM, and executes remediation only through a governed, cryptographically audited path.
4independent evidence lanes behind every diagnosis
5forecast horizons per advisory, minutes to days
3autonomy tiers, from shadow-only to supervised
6,700+automated tests gating the control plane
The problem
Alerts are cheap. Correct, accountable action is not.
Most teams do not lack monitoring. They lack a trustworthy path from a wall of signals to a decision someone is willing to sign their name to.
- 01Signal without meaning. A CPU spike, a 5xx burst and a failed unit are three alerts, not one incident. Correlation is left to the operator, at 3am.
- 02Context lives in people. Runbooks go stale. The engineer who understood the topology left. Onboarding a new system means rebuilding that knowledge by hand.
- 03Automation is untrusted. Teams that could automate remediation often refuse to, because a wrong mutation is worse than a slow human — and because nothing proves afterwards what the automation actually did.
- 04Regulated environments can't send data out. Finance and public-sector operators cannot ship production telemetry to a third-party model API.
How it works
Four evidence lanes, one reasoned advisory.
Every signal enters through a typed lane with its own detection logic. Lanes are independent, so a diagnosis is never built on a single noisy metric.
Hard failure
An OS-level state machine tracks units, mounts and kernel state. Deterministic invariants decide what counts as a real failure before any model is consulted.
Application HTTP
Log-surge probes classify traffic by status family — 5xx, 429, 401 — separating a capacity problem from an abuse problem from an auth regression.
Security
Security incidents are correlated into kill chains across entities and time, then forecast forward, rather than reported as isolated events.
remote agents
read-only collectors on the Linux fleet
alert sources
cluster · application · security
knowledge lane · diagnostic lane
routine observations never enter the alert path, and never invoke the model
retrieval over prior incidents
what happened last time this looked like this
local LLM reasoning
what · who · why · how-to + five-horizon forecast
GOVERNANCE GATE
fail-closed — no audit record, no action
suggest
human approval
execute
outcome feedback
re-evaluated, then learned from
Governance
Autonomy you can switch off, and audit after the fact.
Omni is designed for operators who are accountable for the systems it touches. Every safety property is fail-closed: if the audit path is unavailable, the action does not happen.
Read/write separation
The reasoning path is strictly read-only. Every mutation flows through a single executor with an explicit tool allowlist and namespace isolation. The reasoning layer cannot mutate the cluster even if the model asks it to.
Cryptographic audit trail
Decisions, dispatches, human approvals, trapped mutations and rollbacks are written to a SHA-256 hash chain with Ed25519 signatures — designed against SOX §404 and PCI-DSS v4.0 evidence requirements.
Master kill switch
Automatic execution is disabled by default. Turning it on is a deliberate, per-tenant configuration change, and the effective value is observable at runtime rather than inferred from a document.
| Tier | Behaviour | Intended for |
|---|---|---|
| shadow | Diagnoses and forecasts, takes no action. Advisories are delivered to operators for scoring. | New tenants; building trust and a ground-truth record. |
| minimal | Executes a narrow, reversible set of remediations; everything else escalates to a human. | Tenants with a measured advisory accuracy record. |
| autonomous | Executes within policy, with pre-mutation snapshots and rollback, still fully audited. | Mature, well-understood workloads. |
Engineering
Async-first, event-driven, and able to run entirely on-premise.
Inference runs locally, so production telemetry never leaves the customer's boundary — the requirement that rules out most hosted-model approaches in regulated sectors.
- Control plane
- Python, asyncio end-to-end, FastAPI ingress
- Event backbone
- Kafka — split evidence, action and feedback topics
- Reasoning
- Locally hosted LLM inference, no external model calls
- Retrieval
- Vector search over prior incidents and playbooks
- State
- PostgreSQL for configuration and ledgers; Redis for hot paths
- Targets
- Kubernetes workloads plus agent-instrumented Linux hosts
- Edge agent
- Read-only collector, hardened against command injection and data exfiltration
- Interfaces
- Provider and tenant web portals, HTTP API, chat notifications
Where we are
A working system in a lab environment, moving toward production pilots.
We would rather state this plainly than overclaim. The following reflects what has been verified by running the system, not what is planned.
RUNNINGFull diagnostic pipeline operating continuously against a live Kubernetes cluster and a fleet of instrumented Linux hosts.
RUNNINGMulti-tenant control plane with per-tenant configuration, isolation and enrolled remote agents.
RUNNINGCryptographic audit ledger, human-approval workflow, and provider/tenant portals.
RUNNINGAutomatic system-topology discovery, producing a continuously updated model of each customer environment.
NEXTManaged cloud deployment — GKE for the control plane, Pub/Sub or managed Kafka for the event backbone, Cloud SQL for ledgers, and accelerated inference for tenants who permit hosted models.
NEXTDesign-partner pilots with operators running regulated Kubernetes workloads.
Contact
Talking to infrastructure teams and design partners.
If you operate Kubernetes or a Linux fleet under compliance constraints and the on-call burden is the bottleneck, we would like to hear how you work today.