Rolling and seasonal baselines
Model expected behavior using rolling, robust, seasonal and learned baselines. Show the expected band, observed value and deviation in context so recurring demand patterns can be distinguished from a meaningful change.
Learned baselines, graph intelligence & governed assistance
Recognize unusual behavior in the context of normal demand.
A metric that is normal at one time can be a warning at another. AmrudAi combines seasonal baselines, multivariate and peer analysis, forecasts and graph reasoning with grounded AI assistance so operators can focus on meaningful changes and explain the next decision.
The application experience
Adaptive baselines & anomaly detection
See the complete workspace. Enlarge any image to inspect its panels, navigation and evidence at actual size.
Adaptive Baselines & Anomaly Detection — illustrative product view with sample data. Illustrative product view · Sample data
Compare observed behavior with learned seasonality, quality gates and independently reviewed outcomes. The capability in detail
Make intelligence useful to the operator and accountable to the evidence.
Training windows and sensitivity control
Configure the training window, seasonality, sensitivity, confidence and minimum sample quality for the monitored class. Versioned policy and provenance make it possible to understand why the current baseline applies and how it changed.
Multivariate anomaly detection
Look for related abnormal behavior across multiple metrics instead of assessing every gauge in isolation. A combination of connection pressure, latency and queue growth can guide attention to a developing issue that deserves corroboration.
Peer and cohort comparison
Compare entities with governed groups of relevant peers and check that the evidence is sufficient. Site, service, technology or operating context helps teams distinguish a local outlier from behavior shared across a comparable population.
Forecasts and deterministic what-if analysis
Use trends and statistical models to assess capacity, exhaustion and degradation risk with assumptions and confidence. Controlled scenarios such as demand growth or loss of a resource expose remaining headroom through explicit calculations.
Graph and change intelligence
Use the temporal dependency graph, protocol evidence and change history to prioritize root candidates, propagation patterns and potential impact. Learned graph scores support the causal investigation alongside contradictions and alternative explanations.
Grounded Incident Copilot
Ask questions about the incident, summarize evidence, explain related runbooks and identify useful next tests. Answers connect observed behavior, impact, confidence, evidence and missing information to the authorized operational context.
Local, private and hosted AI choices
Apply approved AI-provider and data-egress profiles to the deployment, including local/private inference and no-LLM operation. Sensitive fields and secrets follow minimization and access policy before any authorized context is supplied.
Model quality, drift and feedback
Track model, prompt, policy and evidence versions with drift and evaluation results. Structured operator feedback and champion/challenger comparison support controlled improvement and rollback without silently changing live policy from a single response.
Resource governance and graceful degradation
Bound AI concurrency, memory, CPU, queues and execution time. Defer expensive recalibration or optional assistance when operational pressure increases, with visible degraded state and deterministic monitoring and investigation remaining available.
From information to action
Use the baseline to guide a defensible investigation.
- 01
Establish expected behavior
Choose the monitored scope, comparable history, seasonality and sensitivity, reviewing the available sample quality.
- 02
Inspect the deviation in context
Compare observed values with the expected band, related metrics, peers and affected service dependencies.
- 03
Verify the operational explanation
Use graph and change evidence, then run the permitted check needed to corroborate or reject the anomaly hypothesis.
- 04
Review the outcome and model quality
Record accepted findings and feedback; use evaluation and controlled version promotion to improve future behavior.
Built around your operating environment
Fit the way your team operates.
Apply statistical, graph and language capabilities to the relevant operational task, with visible provenance and operating controls.
- Rolling, robust, seasonal, learned, multivariate and peer/cohort intelligence.
- Expected bands, sample quality, training window, sensitivity and model version.
- Forecasts, capacity scenarios, graph-assisted RCA and grounded Copilot.
- Local/private or approved hosted providers, data policy, drift evaluation and resource controls.
Connect the workflow
How are anomalies and seasonal baselines evaluated?
How do seasonal baselines help compared with a fixed threshold?
A seasonal baseline describes expected behavior in its recurring operating context, such as the daily or weekly demand pattern configured for the metric. AmrudAi compares the observation with that band and its confidence, while fixed limits can still protect explicit operating boundaries.
Does an anomaly automatically mean something has failed?
No. An anomaly is a candidate that merits context. Telemetry quality, related metrics, service dependencies, changes and verification determine whether it contributes to a qualified issue or a defensible root-cause hypothesis.
Can the platform operate with AI kept inside our environment?
Yes. The documented deployment choices include approved local or private inference through the AI Gateway, with configured access, egress and resource policy. Deterministic monitoring, graph evidence and investigation also operate in a no-LLM mode.
How do we prevent a model from silently changing operational policy?
Models, prompts and policies are versioned. Feedback becomes reviewable evidence for evaluation; drift monitoring and champion/challenger comparisons support controlled promotion or rollback. A single operator response does not directly rewrite live thresholds or privileged-action authority.
See the workflow in your environment.
Explore the technology, service dependencies and operating requirements that matter to your team.