Vishnu Varma

About

AI/ML engineer (Cognizant) who built a production, real-time credit card fraud detection platform combining deep-learning anomaly detection with an LLM-based explanation layer. Strong focus on regulated deployment: addressed class imbalance and feature drift, and added guardrails (SHAP/structured inputs, fine-tuning on analyst reports, rule-based validation) to keep explanations accurate and compliant. Orchestrated the full pipeline with Airflow + Databricks/Spark and used MLflow/Prometheus plus A/B and shadow deployments for measurable reliability.

Experience

Education

Key Strengths

• Built and deployed real-time credit card fraud detection system with hybrid deep learning + LLM explanation layer
• Handled extreme class imbalance (<1% fraud) using SMOTE and cost-sensitive learning to improve recall without increasing false positives
• Mitigated feature drift with online learning and periodic retraining on recent data
• Implemented compliance-focused LLM guardrails (structured inputs, fine-tuning on analyst reports, rule-based validation) to reduce hallucinations
• Orchestrated end-to-end ML workflows in Apache Airflow across ingestion, Databricks/Spark feature jobs, inference, and retraining
• Improved traceability and failure recovery; reduced debugging time ~30% via Airflow task logging/status tracking
• Production reliability practices: unit/integration tests, continuous evaluation (MLflow/Prometheus), drift monitoring, A/B tests and shadow deployments
• Cross-functional collaboration with compliance/legal; translated explainability and data lineage into business terms and built performance/compliance dashboards