Lumis Data & AI · R&D · research prototype
Diagnosis for data pipelines, ML workflows, and the infrastructure beneath them. Deterministic checks first. A bounded investigator when they fall short. A person in control.
What it understands
Lumis Data & AI models your estate in the terms your engineers already use — so an incident is a broken relationship between real things, not a wall of alerts.
Connectors describe what Lumis may read. On the roadmap: Airflow, Databricks, Snowflake, MLflow, Kafka, OpenLineage.
How a diagnosis happens
Known failures can be matched by checks, with no model call. Only what the checks cannot explain goes to a single bounded investigator, with a budget and read-only tools. Below are three real incidents from the GridCast evaluation, as Lumis reported them.
planning-api-scaled-to-zeroMATCHTerminal: the signature explains the alert on its ownfeature-query-amplificationNO MATCHforecast-service-oom-killedNO MATCHforecast-model-slowdownNO MATCHResolved by the check alone in under a second: 14 evidence queries, no model call. The cause: planning-api had been left at zero replicas after a maintenance window.
Every claim is checked against the evidence collected. If it does not hold: “insufficient evidence”.
A structured report reaches an engineer. Conclusion: supported_diagnosis. Correct in all 7 runs: 5 rules-only, 2 Lumis.
Real runs from the GridCast evaluation (4 October 2026; one synthetic estate, one model).Scenario J report and transcript The full study
What you get back
Every investigation ends in structure: the explanations considered, the evidence for and against each, and what is still unknown. Auditable artefacts are stored — not a model's stream of thought.
H1Release 1.7.0 switched lag features to minute resolutionsupportedcheckplanning-api scaled to zeroruled outcheckFeature builds or database auth failingruled outcheckforecast-service OOM or model slowdownruled outLumis Data & AI is a research prototype. Read how it performed on fifteen injected failures, try the open-source SDK, or tell us about a failure class you would like studied.