Articles
Insights, case studies, and lessons from building mission-critical AI systems.
Graph-Based Anti–Money-Laundering Monitoring with Community Detection, Centrality, and Hub Heuristics
A practical methodology for money-flow graphs in digital banking, oriented to layering, structuring, and PPATK-facing review queues
Customer-Level Credit Risk Scoring with Gradient Boosting, Stratified Validation, and APPROVE/REVIEW/DECLINE and Grade A–E Policies for Digital Banking
Credit risk scoring asks how risky a customer's overall posture is—distinct from transaction fraud detection, which asks whether a specific payment event is theft or misuse. We present a research and engineering methodology for customer-level risk that combines (i) aggregation of KYC, account, transaction, login, device, SLIK, fraud-history, and cashflow features, (ii) supervised gradient boosting under class imbalance, (iii) stratified holdout validation with optional stratified K-fold PR-AUC, and (iv) a grade policy mapping continuous scores to APPROVE / REVIEW / DECLINE. Labels (is_default) are constructed from loan defaults, SLIK write-off, and severe DPD hits, or severe delinquency / repayment stress. Primary metrics are PR-AUC and ROC-AUC; Precision, Recall, and F1 support operational tuning. Early-warning rule overlays accompany scores for risk and fraud review teams. The protocol is designed for reproducible experimentation on synthetic or institutional master data and progressive deployment via scoring APIs and dashboards.
Customer-Level Behavioural Risk Scoring with Gradient Boosting, Stratified Validation, and LOW/MEDIUM/HIGH Bucket Policies for Digital Banking
Behavioural risk scoring asks how risky a customer's overall posture is—distinct from transaction fraud detection, which asks whether a specific payment event is theft or misuse. We present a research and engineering methodology for customer-level risk that combines (i) aggregation of KYC, account, transaction, login, device, watchlist, fraud-history, and cashflow features, (ii) supervised gradient boosting under class imbalance, (iii) stratified holdout validation with optional stratified K-fold PR-AUC, and (iv) a bucket policy mapping continuous scores to LOW / MEDIUM / HIGH. Labels (is_high_risk) are constructed from KYC HIGH risk, watchlist hits, or extreme behavioural heuristics. Primary metrics are PR-AUC and ROC-AUC; Precision, Recall, and F1 support operational tuning. Early-warning rule overlays accompany scores for risk and fraud review teams. The protocol is designed for reproducible experimentation on synthetic or institutional master data and progressive deployment via scoring APIs and dashboards.
Transaction-Level Fraud Detection with Gradient Boosting, Chronological Validation, and Decision-Policy Cutoffs for Digital Banking
Payment fraud in digital banking is rare, adversarial, and temporally drifting. We present a practical research and engineering methodology for transaction-level fraud detection that combines (i) multi-source feature engineering from master banking tables, (ii) supervised gradient boosting under class imbalance, (iii) chronological holdout validation to reduce temporal leakage, and (iv) a decision-policy layer that converts continuous fraud scores into operational actions. The unit of analysis is a single transaction. Labels are derived from curated fraud labels and disputes, optionally augmented with weak supervision when positives are scarce. Primary research metrics are Precision–Recall AUC (PR-AUC) and ROC-AUC; thresholded Precision, Recall, and F1 support operational tuning. Reason codes overlay model scores to support analyst investigation. The methodology is designed for reproducible experimentation on synthetic or institutional master data and for progressive deployment into fraud operations workflows.