Graph Neural Networks for Recommendations, Fraud and Connected Data

I build graph machine learning systems for the cases where the signal lives in the connections, not in isolated rows. When your users, accounts, devices, payments or products form a dense web of relationships, a flat table throws away the most valuable feature you have — who is linked to whom. I design and train graph neural networks that read those links directly, so the model reasons about neighborhoods and paths instead of pretending every record stands alone.

My work covers the full pipeline. I start with graph construction: turning raw logs, transactions and catalogs into clean nodes and typed edges, choosing what a connection actually means and how to weight it. From there I build node and edge prediction models, learn embeddings that place similar entities close together, and tune the architecture — GraphSAGE, GAT, message passing and heterogeneous graphs — to fit your data rather than a textbook example. I care as much about honest evaluation as about accuracy, so I set up leakage-safe splits, temporal validation and baselines that tell you whether the graph is really adding value.

Deployment is part of the deal, not an afterthought. I package models for batch scoring or low-latency online inference, handle neighbor sampling for large graphs, and document everything so your team can retrain and extend the system later. Typical outcomes include recommendation engines that understand shared context, fraud and abuse detection that spots rings instead of single bad actors, and enrichment layers that feed graph embeddings into your existing models. If you have relationship-rich data and a suspicion that the connections carry the answer, I can prove it and ship it.