yashumani/drill-down-anamoly

Purpose and runtime

Drill-down Anomaly Lab is a client-side React/TypeScript research prototype for investigating tabular-data anomalies across all eligible dimensions rather than following a predefined drill hierarchy. The browser app accepts CSV/JSON or deterministic demo data, derives anomaly drivers and interactions, and renders an investigation tree.

Entry points are src/main.tsx (React bootstrap) and src/App.tsx (application state and orchestration). Static deployment artifacts are published through .github/workflows/pages.yml.

Data ingestion and schema boundary

src/lib/io.ts parses CSV and JSON into tabular row objects. src/data/sampleData.ts supplies the built-in synthetic dataset for validation and demonstration.

src/lib/profile.ts profiles columns and classifies numeric measures, categorical dimensions, dates, booleans, and identifier-like/high-cardinality fields. This inferred schema is the boundary between arbitrary input data and the analytical engine; only eligible dimensions and selected numeric measures enter scoring.

Shared row, profile, investigation, and visualization contracts are centralized in src/types.ts.

Investigation state and cohort control

src/App.tsx owns the selected Actual measure, optional Expected/Target measure, dataset/profile state, and the active drill path. Each drill selection is treated as an added cohort filter, not as navigation within a static dimension tree.

A new cohort triggers a full re-analysis of the filtered rows. The resulting path records investigation history and allows the UI to compare the current population with selected segment populations.

Anomaly analysis engine

src/lib/anomaly.ts is the core analytical stage. For each cohort it computes row residuals (actual - expected, or a cohort-mean baseline), independently groups all eligible dimensions, and ranks dimensions using interpretable impact, surprise, variance concentration, and support signals.

It then ranks category values by variance contribution and searches pruned two- and three-way interactions among leading dimensions/categories. This is intentionally browser-scale computation; enterprise-scale group-bys, scoring, interaction search, authorization, and row evidence should move behind the same UI-facing contract.

Chat and analytical assistance

src/lib/chatEngine.ts provides the local chat/investigation assistant behavior used by src/components/ChatPanel.tsx. It operates on the app’s current analytical context rather than representing a documented external AI service integration.

The chat panel is therefore a presentation/control surface over client-side investigation state, alongside visual drill actions.

Visualization and user interaction

src/components/Visuals.tsx renders the investigation tree, all-dimension driver landscape, category contributions, interaction segments, KPI summaries, and audit-oriented views. Selecting a category or interaction feeds a new cohort filter back to src/App.tsx.

src/components/EChart.tsx is the reusable chart adapter, isolating chart-library lifecycle concerns from analytical and state-management logic. src/styles.css provides the application styling layer.

Deployment and documentation

The app is a static frontend built with the npm scripts in package.json; no server, database, or external API is represented in the repository. GitHub Pages deployment is configured at .github/workflows/pages.yml.

Research rationale and implementation-oriented architecture notes live in docs/all-dimension-investigation-tree.md and docs/prototype-architecture.md; the generated/static site surface is under docs/ and index.html.