Expanded an existing AI agent from a basic prototype into a more product-oriented application. Added multi-model testing across OpenAI and Claude to compare response quality, performance, and operating cost, and scoped a usage dashboard covering token consumption, latency, and estimated spend.

Built a second, separate RAG pipeline on AWS: a Bedrock Knowledge Base with S3 Vectors for retrieval, Claude Sonnet 4.6 for grounded generation, and Titan embeddings, backed by Lambda, API Gateway, and DynamoDB for per-session conversation history, all provisioned through Terraform. Built to cloud security fundamentals from the start: least-privilege IAM roles, blocked public access on the source bucket, and scoped service-to-service trust relationships. Verified end-to-end via direct API calls, not just console testing, and confirmed the system correctly declines out-of-scope questions rather than hallucinating. Infrastructure was torn down immediately after testing to avoid ongoing charges.

Applied the same supervised-learning and imbalance-handling lens to two different rare-outcome business problems: telecom customer churn (roughly 26.5% positive class) and credit card fraud (roughly 0.17% positive class). Compared Logistic Regression and XGBoost across class weighting, SMOTE, undersampling, and threshold tuning for both. For fraud specifically, threshold-tuned XGBoost was selected as the practical business-balanced model: recall improved from 75.5% to 85.7% (missed fraud cases dropped from 12 to 7), trading precision down to 71.2% (false positives rose from 1 to 17), a deliberate, quantified trade-off rather than an unqualified win.