Triage infrastructure signals
Interpret technical logs and alerts with a tokenizer designed around operational terms and telemetry.
Infrastructure intelligence model
Nova
Operational intelligence.
Inside your environment.
Turn infrastructure logs and alerts into focused triage, agent routing, and relevant runbook recommendations. Nova brings specialised model intelligence to private infrastructure workflows.
Purpose-built intelligence
Help your team find the next useful step.
Route work to specialists
Select a relevant agent domain or escalate when the available evidence does not support a confident route.
Retrieve relevant runbooks
Match incidents and proposed actions through semantic representations, then apply action checks.
Model architecture
One backbone. Five specialised heads.
A transformer backbone with contract-lattice working states connects language diagnostics, agent routing, execution gating, confidence scoring, and contrastive retrieval.
Private deployment
Run locally on CPU or GPU, with PyTorch, SafeTensors, LibTorch, and ONNX deployment formats.
Operational interfaces
Use the triage and action-judge interfaces to connect Nova with your own tools and approval workflows.
Action review
Combine destructive-pattern checks, incident/action relevance, and model gate scores to approve, request peer review, reject, or abstain.
Technical specification
- Backbone
- 12 transformer layers; hidden width 576
- Working states
- 12 contract-lattice slots; 3 routing steps
- Tokenizer
- 32,881 BPE vocabulary; 20 telemetry token classes
- Input window
- 160 tokens for focused log and alert inputs
- Service interfaces
POST /v1/triageandPOST /v1/judge
Engineering evaluation
Inspect the reported measurements.
Internal evaluation: 100 incidents split equally between Kubernetes and VMware, with a 300-runbook retrieval library.
92% top-1 retrieval
98% top-3 retrieval and 0.9467 mean reciprocal rank in the reported runbook evaluation.
271.14 ms mean CPU triage
262.93 ms median and 390.51 ms p95 for the reported incident triage test.
Separate action evaluation
The reported CPU action-judge mean was 689.99 ms. Triage time and action-review time measure different operations.
Source: DevHeal Nova technical evaluation. Results describe that test configuration; deployment performance depends on hardware and workload.
Your workflow. Your first autonomous outcome.
See autonomy in action.
Start with a free pilot.
Curious how autonomous behaviour could work for your business? Bring a real challenge. We’ll help you define a focused pilot, explore the right product, and agree what success looks like in your environment.
Request your free pilot →
No guesswork. Start with a conversation.