At GSTN, I built and deployed machine learning and graph-analytics systems for large-scale tax fraud detection, helping tax officials turn complex data into actionable investigative leads.
I developed a cluster-based anomaly detection model across approximately 90 lakh GSTINs using SQL, PCA, Apriori, Agglomerative clustering, and K-Means. The organization-wide deployment flagged high-risk taxpayer behavior.
I also designed Neo4j graph traversal algorithms to rank transaction chains and produce monthly high-risk GSTIN lists for field teams, contributing to the discovery of a ₹102 crore fraud case in Maharashtra. I built an HSN code prediction web application using a CBOW-based neural network and integrated real-time text auto-completion into internal GST tools.
Alongside production analytics, I've built a GNN fraud-detection model with PyTorch Geometric and applied causal inference methods with DoWhy. I communicate findings through stakeholder-facing dashboards and reports that support faster decisions.
