← Selected systems

CASE STUDY / 02

Real-time fraud intelligence

Use language signals to help identify fraudulent banking transactions.

Based on the public résumé. Architecture is a conceptual walkthrough; client implementation details and private evaluation data are not published.

Responsibility

The résumé describes developing a banking-client fraud detection system using BERT and XGBoost during the Cognizant role.

Architecture at a glance

  1. Transaction context
  2. BERT signals
  3. XGBoost scoring
  4. Risk review

Reported outcome

$2M+ reported annual savings

Annual savings is a business outcome reported in the PDF alongside a 40% reduction in fraudulent transactions; it is not a model accuracy score.

The loss baseline, annualization method, review costs and attribution method are not published. This is a reported system outcome, not a claim of independently verified personal savings.

Read the source résumé (PDF)
Technical deep dive

Design decision

Pair language representations with a decision model so unstructured context can inform risk scoring.

Alternatives and tradeoffs

Rules offer straightforward explanations but need maintenance as patterns change. A learned model needs threshold calibration and monitoring for false positives.

Engineering takeaway

Assess false-positive cost alongside detected fraud. Review capacity and changing transaction patterns matter as much as offline model scores.