Sagarika Dubbaka

Applied Bayesian conversion-rate estimation using Beta models and SQL Server/MySQL data to optimize Amazon Ads bids, evaluating Top-of-Search experiments and SKU/bundle performance to guide spend allocation across three e-commerce accounts. Automated cost of goods and operational expenses invoice processing using Azure Databricks, Azure Document Intelligence, and Power Automate, reducing manual workload by 70% and improving expense visibility for finance. Built Power BI and channel dashboards to present advertising, organic sales, and shipping insights, accelerating cross-functional campaign and operational decisions.

Architected a multi-stage anomaly detection pipeline (whitelist/statistical/unsupervised) for time-series DNS data to reduce false positives, using FFT to filter noise and reduce computation by 60%. Developed an unsupervised detection layer with an LSTM autoencoder (TensorFlow) using TF-IDF vectorization over 10-minute traffic windows, flagging previously unseen low-and-slow threats via reconstruction-error (MSE). Validated end-to-end logic with a malware traffic simulator injecting 10% synthetic attack patterns, demonstrating reliable weighted-score detection behavior to support production planning.

Engineered a customer-facing Virtual AI Assistant using LangChain orchestration, semantic routing, and RAG to automate initial screening and reduce manual support workload by 80% for GTM and sales teams. Deployed a cost-optimized GenAI pipeline on Azure (FastAPI/Docker) using Hugging Face embeddings and Qdrant to resolve 50% of queries locally, routing to Azure OpenAI only for complex edge cases.