REVIEWED LEARNING · EXPERIMENTAL
Human-Reviewed Database Model Learning
Build a reviewable feedback loop around supported numeric PostgreSQL and MySQL observations. Capture, human labels and traceable dataset versions let a team evaluate a candidate without granting it production control.
Guided pilot scope · Reviewed 2026-10-16. No production AI diagnosis, automatic infrastructure/DDL remediation or compliance certification is offered.
Capture only eligible observations
An administrator explicitly enables collection. Eligible records require target identity, complete connection/blocked-session measurements and a recent matching collector-host CPU observation. Incomplete measurements and Redis are skipped.
Collector-host CPU does not establish database-host CPU. Review the source relationship before labeling an observation.
Review the evidence and freeze a dataset
Label normal behavior or concurrent connection saturation, lock contention and compute pressure. Group observations from the same event, correct suggestions and export reviewed records as a frozen checksum-versioned dataset.
Exports omit SQL, hostnames, credentials and reviewer identities. Numeric operational features, timestamps and pseudonymous site/episode identifiers remain; review data-sharing requirements.
Evaluate a candidate before trusting it
The offline Python trainer uses partition and site holdouts, minimum data requirements and episode separation. A candidate may be imported for shadow comparisons; uncertain and near-threshold predictions abstain.
Collection does not automatically train or activate a model. Import validation checks artifact structure and dataset association, not the truth of every training claim.
Keep production decisions with people
This is a Suite 1 advisory workflow, not active AI across all nine suites. Existing experimental suite baselines were trained on synthetic scenarios and remain inactive. Production AI diagnosis, automatic fixes and DDL execution are unavailable.
Collection can be disabled and retained learning records deleted. Previously downloaded datasets are outside that deletion operation. Synthetic scores do not establish diagnostic accuracy for a customer workload.
Frequently asked questions
Are nine AI models diagnosing customer infrastructure now?
No. Experimental synthetic baselines remain inactive. The supported reviewed-data workflow is Suite 1 only and imported candidates stay in advisory shadow mode.
Does opt-in automatically trigger training?
No. Capture, human review, dataset export, offline training and candidate import are separate steps.
Can we disable collection?
An authorized administrator can disable collection and delete retained learning records. Previously exported copies must be handled separately.
Evaluate your own workflow
Start with supported sample or staging infrastructure and agree on a success measure. Contact the operator to arrange a pilot. This static website does not create accounts or send invitations automatically.
Request a guided pilot →