Multimodal Drift Detectives

Dashboards and models often fail silently when data drifts: null spikes, schema changes, or category shifts make KPIs unreliable. Traditional drift tools only flag what changed, leaving teams guessing why. The Multimodal Data Drift Detector combines statistical drift detection with context ingestion from release notes and dashboard screenshots. It detects null %, numeric drift (KS/mean shift), categorical drift (PSI), and outliers, then ingests PDFs and screenshots to capture signals like “currency switched USD→EUR” or “API v2 deployed.” The system generates a plain-English narrative with actionable next steps such as validating ETL, backfilling data, or updating contracts. A simulated chat integration enables instant alerts and team responses. An optional sandbox supports threshold tuning, cost-sensitive metrics, and reproducibility with downloadable configs. This helps teams move beyond detection to diagnosis and action, reducing downtime and keeping metrics trustworthy.