Elysia Tools
Navigation
Workflow Playbook
Profile tabular data, isolate duplicates, missingness, referential breaks, outliers, and time-series anomalies, and turn findings into defensible quality evidence.
Hubs
A quality investigation is more credible when the first artifact is a dataset profile. Capture row counts, column completeness, duplicate pressure, value ranges, and source-to-source diffs before cleaning anything. That baseline lets you explain whether the issue is new, localized, or already present upstream.
Duplicates, broken foreign keys, and out-of-bound values are usually deterministic defects. Filter those records into explicit review buckets first. After structural failures are separated, evaluate distribution behavior with outlier detection, box plots, z-scores, and robust averages so you can tell whether a spike is data corruption, a one-off business event, or a persistent shift.
Nulls are not just blanks; they may signal pipeline outages, delayed submissions, or optional fields. For ordered measurements, compare missing segments with temporal anomaly detection before filling anything. If you interpolate, keep the original series, the repair rule, and the affected intervals so downstream users can decide whether repaired values are fit for forecasting, dashboards, or only exploratory analysis.
The final output should name the affected rows, the violated rules, the anomalous windows, and the recommended action for each class of finding. Good evidence separates confirmed defects from tolerated exceptions and makes the remediation path reproducible.
Workflow playbook
Measure completeness, column shape, suspicious value patterns, and comparison deltas before making judgments about quality drift.
Isolate duplicate rows, filter suspect slices, and validate boundaries and foreign keys so obvious rule breaks are not mixed with statistical anomalies.
Use outlier scoring and robust summary statistics to determine whether extreme values represent bad data, rare real events, or skewed but acceptable behavior.
Review gaps and temporal shocks, interpolate only where justified, and capture before-and-after evidence for every adjusted series.
Agree on acceptable blanks, seasonal spikes, duplicate tolerances, and valid boundary ranges before labeling anomalies.
Use interpolation only where sequence continuity is expected; otherwise preserve null evidence for manual review.
Use record-level detectors for duplicates and foreign keys, then compare aggregate distributions to confirm systemic drift.