docs: improve enterprise code documentation, formatting, and 10-agent Copilot architecture details

fix: implement pure Python NeighborLoader fallback and weights_only=False in eval script

feat: add hybrid GNN/ML models and improved investigation UI

refactor: implement network community overview sampling, fix ruff unused import, and untrack ruff cache

ml: retrain XGBoost from scratch + restore original int(proba*99) scoring

fix: calibrate XGBoost risk scores & classify alert queue by tier

Merge origin/chauhan: bring in UI redesign and frontend enhancements

UI: Align alert cards risk badges and colors, make export buttons green, customize status dropdown select, and clean overview CTAs

fix: recompute GNN feature statistics from dataset at runtime, not checkpoint

fix: build SHAP input DataFrame from pre-coerced dict, not chained indexing

fix: GNN NaN scores + SHAP tab animation and invisible title

fix: coerce feature columns to float before XGBoost/SHAP, fix alert enrichment skipping score=0

fix: replace shap.TreeExplainer with XGBoost native pred_contribs

fix: build SHAP input as numpy float64 array, not pandas DataFrame

fix: remove simulated XGBoost metrics; harden SHAP against string dtypes

feat: download models from HF Hub at startup; improve GNN trainer

fix: recompute GNN feature statistics from dataset at runtime, not checkpoint

feat: integrate case management with Supabase authentication and storage

Improve investigation GNN and metric display

Stabilize investigation SHAP handling

Harden investigation numeric coercion

feat: decouple frontend and fix investigation 500 error

feat: Implement caching for API responses and SHAP explanations to improve performance and reduce redundant computations.

Fix investigation view rendering, fix STR PDF generation (unicode/limit), and add graph checkpoint for fast server restart

changed

Implement ML fraud detection pipeline and interactive graph visualiser

Initial commit: Fund Flow Tracker AML Dashboard