Shubham Singh

Machine Learning Engineer

• Architected a Python-based LLM-powered internal chatbot using LangChain and RAG — agentic workflows with semantic search over a Pinecone vector store, now serving 1,200+ employees at 88% resolution accuracy while meeting banking-compliance standards.
• Built end-to-end Snowflake preprocessing pipelines processing 500K+ daily transactions, integrating versioned feature stores for batch training and real-time inference across fraud and credit-risk models.
• Developed supervised ML and anomaly-detection models (scikit-learn, XGBoost) for transaction categorization and fraud detection, reducing manual reviews and false-positive alerts by ~15%.
• Built NLP classification systems with TensorFlow and BERT on AWS ECS handling 50K+ monthly requests at 82% routing accuracy, cutting average response time by 40%.
• Designed ML training and monitoring pipelines via AWS Step Functions, Apache Airflow, and MLflow — automated retraining, experiment tracking, and drift detection at 85%+ production accuracy.
• Implemented A/B testing frameworks and shadow deployments with SHAP-based explainability to satisfy model-fairness requirements and regulatory audits for credit decision systems.
• Containerized services with Docker on AWS EKS with horizontal pod autoscaling and blue-green rollouts to hold SLA targets under variable production load.
• Established pytest frameworks achieving 92% code coverage across unit, integration, and E2E tests for inference APIs and data pipelines.

Machine Learning Engineer

• Designed and productionized transformer-based recommendation systems in PyTorch, increasing user engagement by 23% — validated through A/B tests and statistically significant hypothesis testing tracked in MLflow.
• Built real-time collaborative and content-based filtering pipelines on streaming user events, delivering personalized recommendations with sub-200ms p99 response times.
• Containerized models with Docker, orchestrated on Kubernetes, and served REST APIs handling 5,000+ concurrent requests within uptime SLAs.
• Implemented model-drift detection and automated retraining with MLflow and scikit-learn; ran hyperparameter optimization with Optuna for an 18% reduction in inference latency.
• Engineered Pandas/NumPy feature pipelines for clickstream, purchase, and session data; stored and monitored KPIs (AUC, precision, recall) in PostgreSQL with automated retraining alerts.
• Applied transfer learning to fine-tune pre-trained models on domain-specific catalog data — reducing cold-start issues for new items and improving recommendation diversity.
• Contributed ML-driven features generating an estimated $800K in additional annual revenue; collaborated cross-functionally in Agile sprints.