Chinmay
Sonar.

Problem Model long-range account behavior to identify fraud patterns that are difficult to capture from isolated transactions.

Approach Led development of an encoder-based transformer for transaction sequence modeling and account-holder representation learning, trained at scale using distributed GPU infrastructure.

Impact Trained on 200M+ transaction sequences and delivered a 25% uplift in fraud dollars captured over the production baseline, representing $2M+ in incremental monthly fraud capture.

Problem Identify fraud and identity-risk patterns that are invisible when accounts and transactions are modeled independently.

Approach Led development of graph-based identity representations and proximity embeddings that capture structural relationships between accounts and known fraud networks.

Impact Built representations across 1B+ accounts and 5.4B+ transactions, improving downstream first-party fraud models by 8–15%.

Problem Detect mule accounts and scams across financial institutions and markets with different fraud patterns and data distributions.

Approach Developed mule and scam detection models and extended market-specific modeling into a global framework designed to generalize across regions.

Impact Achieved 70% mule detection with a two-month lead over police reporting in the initial deployment; subsequent global modeling operated across 50M+ monthly transactions.

Problem Reduce manual effort and knowledge loss across complex, repeatable data-science workflows.

Approach Designed an agentic ML system that translates modeling requests into an end-to-end nine-stage workflow spanning problem framing, feature engineering, experimentation, evaluation, and delivery.

Impact Enables validated experiments, features, and expert corrections to compound across projects rather than reset between workflows; currently piloted within Visa.